| Algorithm1Training Algorithm |
| 1: procedure TRAINING(Categorical Distribution D,with M nodes and N categories) Let ian integer from O to M-1 |
| 2: 34 | |
| for kpretrain steps do |
| 5: | x~D |
| 6: | c ~ Ber(σ(γ)) |
| 7: 8: | L = -log P(xlc) > Compute log-probability of data given config |
| 0slow ←Adam(0slow,VL) |
| 9: for kintervention steps do | |
| 10: | I_N←randint(O，M-1) |
| 11: | Dint := D with intervention on node I_N |
| 12: | if predicting intervention then |
| 13: | Li←O∀i |
| 14: | for Kpredict steps do |
| 15: | x~Dint |
| 16: | c ~ Ber(σ(γ)) |
| 17: 18: | Li←Li+-logPi(xlci;Oslow)∀i |
| >Accumulate NLL for every node i separately I_N←argmax(Li) |
| |
| 19: 20: | gammagrads，logregrets =[]，[] >Transfer Episode Adaptation Loop |
| 21: | for Kepisode steps do x~Dint |
| 22: | gammagrad，logregret = 0，0 |
| 23: | for kcfg steps do |
| 24: | c ~ Ber(σ(γ)) |
| 25: | Li=-log Pi(xlci;0slow）∀i |
| 26: | gammagrad += σ(γ)-c |
| 27: | logregret += ∑ Li |
| 28: | I_N |
| 29: | gammagrads.append(gammagrad) logregrets.append(logregret) |
| |
| 30: | J ←入MaxEnt LMaxEnt(γ)+ 入sparse LSparse(γ)+ 入DAG LDAG(γ) |
| 31: | Vγ←VγJ+∑gammagradskijlogregrets.softmax(0) ki |
| 32: | k γ ← Adam(γ,∀γ) |
+
+# A.3 PRELIMINARIES
+
+Interventions In a purely-observational setting, it is known that causal graphs can be distinguished only up to a Markov equivalence class. In order to identify the true causal graph intervention data is needed (Eberhardt et al., 2012). Several types of common interventions may be available (Eaton & Murphy, 2007a). These are: No intervention: only observational data is obtained from the ground truth causal model. Hard/perfect: the value of a single or several variables is fixed and then ancestral sampling is performed on the other variables. Soft/imperfect: the conditional distribution of the variable on which the intervention is performed is changed. Uncertain: the learner is not sure of which variable exactly the intervention affected directly. Here we make use of soft interventions for several reasons: First, they include hard interventions as a limiting case and hence are more general. Second, in many real-world scenarios, it is more difficult to perform a hard intervention compared to a soft one. We also deal with a special case of uncertain interventions, where the variable selected for intervention is random and unknown. We call these unidentified or unknown interventions.
+
+Causal sufficiency The inability to distinguish which causal graph, within a Markov equivalence class, is the correct one in the purely-observational setting is called the identifiability problem. In our setting, all variables are observed (there are no latent confounders) and all interventions are random and independent. Hence, within our setting the true causal graph is always identifiable in principle (Eberhardt et al., 2012; Heinze-Deml et al., 2018). We consider here situations where a single variable is randomly selected and intervened upon with a soft or imprecise intervention, its identity is unknown and must be inferred.
+
+# A.4 EXPERIMENTAL SETUP
+
+For all datasets, the weight parameters for the learned model is initialized randomly. In order to not bias the structural parameters, all $\gamma$ is initialized to 0.5 in the beginning of training.
+
+# A.5 SYNTHETIC DATA
+
+SCM with $n$ variables is modeled by $n$ feedforward neural networks (MLPs) as described in section 3.1. For simplicity, we assume use an acyclic causal graph such that we could easily sample from it. Hence, given any pair of random variables $A$ and $B$ , either $A B$ , $B A$ or $A$ and $B$ are independent.
+
+The MLP representing the ground-truth SCM has its weights $\theta$ initialized use orthogonal initialization with gain 2.5 and the biases are initialized using a uniform initialization between $- 1 . 1$ and 1.1, which was empirically found to yield "interesting" yet learnable random SCMs.
+
+
+Figure 7: Left: Every possible 3-variable connected DAG. Right: Cross entropy for edge probability between learned and ground-truth SCM for all 3-variable SCMs.
+
+# A.5.1 BNLEARN DATA REPOSITORY
+
+The repo contains many datasets with various sizes and structures modeling different variables. We evaluate our model on 3 of the datasets in the repo, namely the Earthquake (Korb & Nicholson, 2010), Cancer (Korb & Nicholson, 2010) and Asia (Lauritzen & Spiegelhalter, 1988) datasets. The ground-truth model structure for the Cancer (Korb & Nicholson, 2010) and Earthquake (Korb & Nicholson, 2010) datasets are shown in Figure 8. Note that even though the structure for the 2 datasets seems to be the same, the conditional probability tables (CPTs) for these 2 datasets are very different and hence results in different structured causal models (SCMs) for the 2 datasets.
+
+
+Figure 8: Left: Ground Truth SCM for Cancer. Middle: Groundtruth SCM for Earthquake. Right: Groundtruth SCM for Asia.
+
+# A.5.2 TRAINING GROUND-TRUTH DATA GENERATOR
+
+Because a CPT is capable of representing any distribution, and MLPs are strictly less powerful in this respect, it may not be possible to learn perfectly the distribution with our MLP learner model. We therefore train a near-ground-truth MLP to replicate as closely as possible the CPT’s probability table, and then use this trained MLP are the ground-truth SCM data generator.
+
+Training is by 1000 iterations of full-batch gradient descent with learning rate 0.001 and momentum 0.9, with all possible parent values masked with the ground-truth vector $\gamma _ { i }$ . The objective is to minimize mean squared error between the MLP’s logits and the log-probability as drawn from the CPT. If the CPT contains a zero, it is approximated by a logit of $- 1 0 0$ .
+
+Given that some of the CPTs contain very unlikely events, we have found it necessary to add a temperature parameter in order to make them more frequent. The near-ground-truth MLP model’s logit outputs are divided by the temperature before being used for sampling. Temperatures above 1 result in more uniform distributions for all causal variables; Temperatures below 1 result in less uniform, sharper distributions that peak around the most likely value. We find empirically that a temperature of about 2 is required for our BnLearn benchmarks.
+
+# A.6 EFFECT OF SPARSITY
+
+We use a $L 1$ regularizer on the structure parameters $\gamma$ to encourage a sparse representation of edges in the causal graph. In order to better understand the effect of the $L 1$ regularizer, we conducted ablation studies on the $L 1$ regularizer. It seems that the regularizer has an small effect on rate of converges and that the model converges faster with the regularizer. This is shown in Figure 9
+
+
+Figure 9: Effect of Sparsity: On 5 variable, 6 variable and 8 variable Nodes
+
+# A.7 EFFECT OF TEMPERATURE
+
+As noted in section A.5.1, we have introduced a temperature hyperparameter in order to encourage the groundtruth model to generate some very rare events in the conditional probability tables (CPTs) more frequently. We run ablation studies to understand the importance of the temperature term. A temperature of 1 corresponds to no changes to the underlying CPTs. As shown in Figure 12, for the Cancer (Korb & Nicholson, 2010) dataset, a temperature of 2 improves the accuracy of causal graph recovery.
+
+
+Figure 10: Cross entropy for edge probability between learned and ground-truth SCM for Cancer at varying temperatures.
+
+
+Figure 11: Cross entropy for edge probability between learned and ground-truth SCM. Left: The Earthquake dataset with 6 variables. Right: The Asia dataset with 8 variables
+
+
+Figure 12: Left: SCM for cross5 graph. Right: Cross entropy for edge probability between learned and ground-truth SCM for chain5 and cross5, comparing the learning process with prediction of edges to ground-truth intervention
\ No newline at end of file
diff --git a/parse/train/H1gN6kSFwS/H1gN6kSFwS_content_list.json b/parse/train/H1gN6kSFwS/H1gN6kSFwS_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..f5a9ad0e6e043eaeccdc02379df47e43f8f0a1cf
--- /dev/null
+++ b/parse/train/H1gN6kSFwS/H1gN6kSFwS_content_list.json
@@ -0,0 +1,2050 @@
+[
+ {
+ "type": "text",
+ "text": "LEARNING NEURAL CAUSAL MODELS FROM UNKNOWN INTERVENTIONS ",
+ "text_level": 1,
+ "bbox": [
+ 230,
+ 98,
+ 767,
+ 146
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Anonymous authors Paper under double-blind review ",
+ "bbox": [
+ 183,
+ 170,
+ 400,
+ 198
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "ABSTRACT ",
+ "text_level": 1,
+ "bbox": [
+ 454,
+ 224,
+ 544,
+ 239
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Meta-learning over a set of distributions can be interpreted as learning different types of parameters corresponding to short-term vs long-term aspects of the mechanisms underlying the generation of data. These are respectively captured by quickly-changing parameters and slowly-changing meta-parameters. We present a new framework for meta-learning causal models where the relationship between each variable and its parents is modeled by a neural network, modulated by structural meta-parameters which capture the overall topology of a directed graphical model. Our approach avoids a discrete search over models in favour of a continuous optimization procedure. We study a setting where interventional distributions are induced as a result of a random intervention on a single unknown variable of an unknown ground truth causal model, and the observations arising after such an intervention constitute one meta-example. To disentangle the slow-changing aspects of each conditional from the fast-changing adaptations to each intervention, we parametrize the neural network into fast parameters and slow meta-parameters. We introduce a meta-learning objective that favours solutions robust to frequent but sparse interventional distribution change, and which generalize well to previously unseen interventions. Optimizing this objective is shown experimentally to recover the structure of the causal graph. Finally, we find that when the learner is unaware of the intervention variable, it is able to infer that information, improving results further and focusing the parameter and meta-parameter updates where needed. ",
+ "bbox": [
+ 233,
+ 257,
+ 766,
+ 534
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 INTRODUCTION ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 565,
+ 336,
+ 580
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "A major challenge of contemporary deep learning is to generalize well outside the assumptions of independent and identically distributed data, when we care about generalization or fast adaptation to distributions other than the main training distribution. For this purpose, we propose an approach that starts by distinguishing between: (a) an underlying causal model, (b) observational distributions derived from it, an (c) interventional distributions arising from interventions upon its variables, whether they be known or unknown. Fortunately, these distinctions can be addressed by the paradigm of Structural Causal Models (SCMs) (Pearl, 1995; Peters et al., 2017) and a wide body of associated literature. Unlike other frameworks using SCMs we also account for interventions performed by agents other than an experimenter. By treating the interventions of other agents as unknown interventions that lead to changes in the underlying data distribution, the present work is a contribution towards the use of a meta-learning for causal model induction. ",
+ "bbox": [
+ 174,
+ 597,
+ 825,
+ 750
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Estimating the underlying causal structure from data is an open and challenging problem (Pearl, 2009; Imbens & Rubin, 2015). A lot of prior work has examined learning causal structure based on observational data (Chickering, 2002; Tsamardinos et al., 2006; Goudet et al., 2017; Hauser & Bühlmann, 2012; Spirtes et al., 2000; Sun et al., 2007; Zhang et al., 2012; Shimizu et al., 2006; Hoyer et al., 2009; Daniusis et al., 2012; Budhathoki & Vreeken, 2017; Kalainathan et al., 2018). However, many real-world datasets have an inherent distributional heterogeneity due to different interventions to the variables composing the model. In these situations interventional approaches are needed (e.g. Heckerman et al., 1995; Cooper & Yoo, 1999; Hauser & Bühlmann, 2012; Peters et al., 2016; Rothenhäusler et al., 2015; Ghassami et al., 2017). Established approaches for causal inference are often either relying on restrictive assumptions or on conditional independence testing, which is hard (Shah & Peters, 2018). Furthermore, most of these approaches either assume full knowledge of the intervention or make strong assumptions about its form (Heinze-Deml et al., 2018). ",
+ "bbox": [
+ 174,
+ 757,
+ 826,
+ 922
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "However, in the real world, interventions are also not always performed by an experimenter. They can be performed by other agents, or by environmental changes in ways that are unknown, or by a naive learner (like a robot) which does not know precisely yet how its low-level actions change high-level causal variables. In this paper, we look at the setting where interventions are unknown, and our goal is to discover causal graphs given unknown-intervention samples. The challenging aspect of this setting is to not only learn the causal graph structure, but also predict the intervention accurately. In this setting, we need to make sure to: (1) avoid an exponential search over all possible DAGs, (2) handle unknown interventions, (3) model the effect of interventions, and (4) model the underlying causal structure. ",
+ "bbox": [
+ 174,
+ 103,
+ 825,
+ 228
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "One possibility for learning a causal structure (through SCM modelling) is to perform many experiments in which one executes interventions. Thus, such interventions modify the effect of the intervened upon variable from its parents in the corresponding DAG, which the model has to quickly adapt to. We can make parallel connections to meta-learning, where the inner loop can be considered as fast adaptation to the distribution change, and outer loop can be considered as learning the stationary meta-parameters of the model. For causal induction, one can consider each distribution which arises as a result of an intervention as a meta-example, and use a meta-learning objective for fast adaptation in response to an intervention. One can think of model parameters as being composed of slow- and fast-changing parameters. The slow parameters are analogous to the meta-parameters in meta-learning and are used for (1) intervention prediction in order to handle the unknown intervention and for (2) modeling the underlying causal structure. On the other hand, the fast parameters are used to model the effect of interventions. An explicit search over the exponentially-growing space of all possible DAGs is avoided by modeling the conditionals of the structural causal model using function approximators, with one neural network per variable. The belief over whether one node $i$ is a direct causal parent of another node $j$ corresponds to a dropout probability for the $i$ -th input of network $j$ (which predicts variable $j$ given its parents). This cheaply represents all $2 ^ { M ^ { 2 } }$ possible model graphs, with the graph search implicitly achieved by learning these dropout probabilities. We thus propose a new method for fast adaptation and learning of neural causal models by framing the problem in a meta-learning setting, similar to (Dasgupta et al., 2019; Bengio et al., 2019). ",
+ "bbox": [
+ 174,
+ 234,
+ 825,
+ 502
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Our contributions Our key contributions can be summarized as follows: ",
+ "bbox": [
+ 174,
+ 508,
+ 651,
+ 523
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "• Handle causal induction to the case where the variable on which a soft intervention took place is not known by the learner, and show that better results can be obtained when the learner attempts to infer that information and uses it to appropriately change parameters and meta-parameters. \n• We bypass the issue of having to optimize over and represent an exponentially large set of discrete causal graphs by learning an efficiently parametrized ensemble of SCMs, \n• Show that our algorithm correctly identifies the causal graph and use the learned graph for generalization to an unseen environment. ",
+ "bbox": [
+ 215,
+ 535,
+ 825,
+ 657
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2 PRELIMINARIES ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 676,
+ 339,
+ 693
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "A Structural Causal Model (SCM) (Peters et al., 2017) over a finite number $M$ of random variables $X _ { i }$ is a set of structural assignments ",
+ "bbox": [
+ 174,
+ 707,
+ 823,
+ 737
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "equation",
+ "img_path": "images/216c3e36b771714960a490774f1f132f5450e31656442971f4f56bc6dd455527.jpg",
+ "text": "$$\nX _ { i } : = f _ { i } ( X _ { p a ( i , C ) } , N _ { i } ) , \\quad \\forall i \\in \\{ 0 , \\dots , M - 1 \\}\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 334,
+ 743,
+ 661,
+ 762
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "where $N _ { i }$ is jointly-independent noise and $p a ( i , C )$ is the set of parents (direct causes) of variable $i$ under configuration $C$ of the SCM directed acyclic graph, i.e., $C \\in \\{ 0 , 1 \\} ^ { M \\times M }$ , with $c _ { i j } = 1$ if node $i$ has node $j$ as a parent (equivalently, $X _ { j } \\in X _ { p a ( i , C ) }$ ; i.e. $X _ { j }$ is a direct cause of $X _ { i }$ ). The $n$ -th power of an adjacency matrix, $C ^ { n }$ , counts the number of length- $n$ walks from node $i$ to node $j$ of the graph in element $c _ { i j }$ . The trace of the $n$ -th power of an adjacency matrix, $\\mathrm { T r } ( C ^ { n } )$ , counts the number of length- $\\mathbf { \\nabla } \\cdot n$ cycles in the graph. Causal structure learning is the recovery of the ground-truth $C$ from observational and interventional studies. ",
+ "bbox": [
+ 174,
+ 767,
+ 825,
+ 866
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Functional and structural meta-parameters Let us consider the simplest SCMs, those with $M = 2$ random variables (lets say random variables $A$ and $B$ ). Only three DAGs exist relating them; they are $A B$ , $B A$ or $A \\bot B$ . These can be represented as the following. If $A$ causes ",
+ "bbox": [
+ 176,
+ 882,
+ 823,
+ 924
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "$B$ , then the SCM between $A$ and $B$ can be represented as $B = f _ { \\theta _ { B } } ( c _ { B A } \\cdot A , \\epsilon _ { B } )$ , where $c _ { B A } = 1$ $\\epsilon _ { B } \\sim N _ { B }$ is a sample of an independent noise factor. If $B$ causes $A$ , then the SCM can be presented as $A = f _ { \\theta _ { A } } ( c _ { A B } \\cdot B , \\epsilon _ { A } )$ , where $c _ { A B } = 1$ , $\\epsilon _ { A } \\sim N _ { A }$ is a sample of an independent noise factor. If $A$ and $B$ are independent, then we have the same equations but we set $c _ { A B } = 0$ and $c _ { B A } = 0$ . ",
+ "bbox": [
+ 174,
+ 103,
+ 825,
+ 160
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "We may now think of learning the structural causal model as learning the two probabilities of $c _ { A B }$ or $c _ { B A }$ being 1 (versus 0), representing our belief in the causal relationship between $A$ and $B$ . We can parametrize these probabilities differentiably using $P ( c _ { A B } = 1 ) ^ { \\bar { } } = \\sigma ( \\gamma _ { A B } )$ and $P ( c _ { B A } = 1 ) = \\sigma ( \\gamma _ { B A } )$ , with $\\gamma$ real numbers and $\\begin{array} { r } { \\sigma ( x ) \\ = \\ \\frac { \\bar { } 1 } { 1 + e ^ { - x } } } \\end{array}$ . We can also parametrize the structural equations $f _ { \\theta _ { A } }$ and $f _ { \\theta _ { B } }$ in a differentiable manner, using conditional probability tables (CPTs) or neural networks. ",
+ "bbox": [
+ 173,
+ 166,
+ 825,
+ 251
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Hence, the problem becomes simultaneously learning the structural meta-parameters $\\gamma$ and the functional meta-parameters $\\theta$ . Functional meta-parameters $\\theta$ may be easily learned by maximum likelihood or a proxy, and using backpropagation. However, the $\\gamma$ are more difficult to infer. Bengio et al. (2019) propose to learn them by using observations of how the distribution changes sparsely, with the transfer generalization (the adaptation rate after the sparse change) being the training objective for $\\gamma$ . Simultaneously inferring both $\\theta$ and $\\gamma$ is still more difficult. An $M$ -variable SCM over random variables $X _ { i }$ , $i \\in \\{ 0 , \\ldots , M - 1 \\}$ can induce a super-exponential number of adjacency matrices $C$ . The super-exponential explosion in the number of potential graph connectivity patterns and the super-exponentially growing storage requirements of their defining conditional probability tables make CPT-based parametrizations of the structural assignments $f _ { i }$ increasingly unwieldy as $M$ scales. As shown below, neural networks with $c _ { i j }$ -masked inputs can provide a more manageable parametrization. For more background about different kinds of intervention we ask the reader to refer Appendix A.3. ",
+ "bbox": [
+ 173,
+ 258,
+ 825,
+ 439
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "3 PROPOSED FRAMEWORK: META LEARNING FOR CAUSAL INDUCTION ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 455,
+ 785,
+ 472
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Our framework disentangles the slow-changing meta-parameters, which reflect the stationary properties discovered by the learner, and the fast-changing parameters, which adapt in response to interventional changes in distribution. We consider two kinds of meta-parameters: the causal graph structure $\\gamma$ and the model’s slow weights $\\theta _ { \\mathrm { s l o w } }$ , along with the meta-learning objective for both of them. We also consider one kind of parameter: the model’s fast weights, $\\theta _ { \\mathrm { f a s t } }$ . We will call $\\theta = \\theta _ { \\mathrm { s l o w } } + \\theta _ { \\mathrm { f a s t } }$ the sum of the slow, stationary meta-parameters and the fast, adaptational parameters. ",
+ "bbox": [
+ 174,
+ 487,
+ 825,
+ 570
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "3.1 TASK DESCRIPTION ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 580,
+ 352,
+ 594
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Our task setup deviates from most common deep learning modeling setups. For the purposes of this work, we restrict ourselves to inference of randomly-generated or manually-provided ground-truth SCMs of $M$ categorical random variables causally related via a DAG. The model is permitted to see (1) data from the original ground-truth model, and (2) data from a modified ground-truth model with a random intervention applied. In our experiments, at most one intervention is concurrently performed. When an intervention is performed, a single node is randomly and uniformly chosen among all $M$ nodes, and its ground-truth distribution soft-intervened upon. The learner model is aware of the samples having come from an intervention distribution, but is not aware of the identity of the intervention node, and so must predict it. Each run of sampling steps under a given intervention is referred to as an episode. The learner, over a large number of episodes, will experience all nodes being intervened upon, and should be able to infer the SCM from these interventions. ",
+ "bbox": [
+ 174,
+ 606,
+ 825,
+ 758
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "3.2 CAUSAL INDUCTION AS AN OPTIMIZATION PROBLEM ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 776,
+ 584,
+ 790
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "We first explain how we mitigate the problem of searching in the super-exponential set of graph structures. If there are $M$ such variables, the strategy of considering all the possible structural graphs as separate hypotheses is not feasible because it would require maintaining $O ( 2 ^ { M ^ { 2 } } )$ models of the data. We note that we can cheaply choose any of the $2 ^ { M ^ { 2 } }$ possible DAGs through suitable independent Bernoulli choices $c _ { i j }$ associated with each edge $i j$ of the causal graph, i.e., sampling all the $c _ { i j }$ ’s independently. Then we only need to learn the $M ^ { 2 }$ coefficients $\\gamma _ { i j }$ , and we implicitly maintain a distribution over the $2 ^ { M ^ { 2 } }$ models corresponding to all the possible draws of $c _ { i j }$ . Note that a slight dependency between the $c _ { i j }$ is induced if we require the causal graph to be acyclic (which allows one to sample the $X$ using ancestral sampling). To enforce that constraint it is not sufficient to require $c _ { i j } c _ { j i } = 0$ (both cannot be 1). We deal with this problem with a regularizer acting on the $\\gamma$ ’s in order to favour acyclic solutions (Zheng et al., 2018). ",
+ "bbox": [
+ 174,
+ 801,
+ 825,
+ 924
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 174,
+ 103,
+ 823,
+ 146
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "In our approach, each random variable’s structural assignment is modeled via $X _ { i } : = f _ { \\theta _ { i } } ( \\bar { c } _ { i 0 } \\times X _ { 0 } , c _ { i 1 } \\times$ $X _ { 1 } , . . . , c _ { i m } \\times X _ { m } , \\epsilon _ { i } )$ , where $f _ { \\theta _ { i } } ( \\dot { ) }$ is a neural network (MLP) with parameters $\\theta _ { i }$ , and $c _ { i j } \\sim \\mathrm { B i n } ( \\mathrm { s i g m o i d } ( \\gamma _ { i j } ) )$ . Through this construction we can frame the causal induction problem as an optimization problem, with $\\theta$ optimized to maximize the likelihood of data under the model but $\\gamma$ optimized with respect to a meta-learning objective arising from changes in distribution because of interventions. There are a few benefits for learning a parametrized ensemble of SCMs. Such an ensemble is analogous to an ensemble of neural nets differing by their binary input dropout masks, which select what variables are used as predictors of another variable. ",
+ "bbox": [
+ 174,
+ 150,
+ 421,
+ 415
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "image",
+ "img_path": "images/d47a155d4b265e10c72c13a432ae0b74490eabb21b453e8a46d108a4b05b1c95.jpg",
+ "image_caption": [
+ "Figure 1: MLP Model Architecture for $M = 3$ , $N = 2$ (fork3) SCM. The model computes the conditional probabilities of ${ \\hat { A } } , { \\hat { B } } , { \\hat { C } }$ given their parents using a stack of three independent MLPs. The MLP input layer uses an adjacency matrix sampled from $\\mathrm { B e r } ( \\sigma ( \\gamma ) )$ as an input mask to force the model to make use only of parent nodes to predict their child node. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 441,
+ 167,
+ 813,
+ 330
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 174,
+ 416,
+ 629,
+ 429
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "3.3 FAST ADAPTATION BY META-LEARNING ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 449,
+ 496,
+ 463
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Fast and slow weights To disentangle an environment’s stable, unchanging properties (the causal structure) from unstable, changing properties (the effects of an intervention), we proposed in $\\ S 3$ to distinguish between the model’s functional meta-parameters $\\theta _ { \\mathrm { s l o w } }$ and parameters $\\theta _ { \\mathrm { f a s t } }$ . The sum of these weights, $\\theta = \\theta _ { \\mathrm { s l o w } } + \\theta _ { \\mathrm { f a s t } }$ , parametrizes the MLPs computing the conditionals $P _ { i } ( X _ { i } | X _ { p a ( i ) } ; \\theta _ { i } )$ The fast weights and the slow weights terminology is drawn from Hinton & Plaut (1987). The construction of $\\theta$ as a sum of initial, slow weights plus zeroed, fast weight that are then allowed to quickly adapt during a transfer episode is due to MAML (Finn et al., 2017). The ability to generalize out-of-distribution by adapting to a transfer distribution can then be measured by the likelihood after adapting the fast weights on transfer data. ",
+ "bbox": [
+ 173,
+ 476,
+ 825,
+ 602
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Since an intervention is generally not persistent from one transfer distribution to another, the model’s functional parameters $( \\theta _ { \\mathrm { f a s t } } )$ are reset after each episode of transfer distribution adaptation. The meta-parameters $( \\theta _ { \\mathrm { { s l o w } } } , \\gamma )$ are preserved, then updated after each episode. Inspired by Bengio et al. (2019), the meta-objective for each meta-example over some intervention distribution $D _ { \\mathrm { i n t } }$ is the following1 \"meta-transfer\" loss: ",
+ "bbox": [
+ 174,
+ 608,
+ 825,
+ 678
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "img_path": "images/068e6a8e648f3a826e74fc8ffaa9b259768beb6cc68fd5adcb940044beeb4554.jpg",
+ "text": "$$\n\\mathcal { R } = - \\mathbb { E } _ { X \\sim D _ { \\mathrm { i n t } } } [ \\log \\mathbb { E } _ { C \\sim \\mathrm { B e r } ( \\gamma ) } [ \\prod _ { i } \\mathcal { L } _ { C , i } ( X ; \\theta _ { \\mathrm { s l o w } } ) ] ]\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 330,
+ 688,
+ 668,
+ 720
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $X$ is an example sampled from the intervention distribution $D _ { \\mathrm { i n t } }$ , $C$ is an adjacency matrix drawn from our belief distribution (parametrized by $\\gamma$ ) about graph structure configurations and ",
+ "bbox": [
+ 171,
+ 728,
+ 825,
+ 758
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "img_path": "images/546c50a94ce4f3eb42178e4df94a5efda19506263a40169ffcd706ef0871bded.jpg",
+ "text": "$$\n\\mathcal { L } _ { C , i } ( X ) = P ( X _ { i } | X _ { p a ( i , C ) } ; \\theta _ { \\mathrm { s l o w } } )\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 385,
+ 767,
+ 612,
+ 786
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "is the likelihood of the $i$ -th variable $X _ { i }$ of the sample $X$ , when predicting it under the configuration $C$ from the set of its putative parents, $X _ { p a ( i , C ) }$ . ",
+ "bbox": [
+ 176,
+ 792,
+ 823,
+ 824
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Structural Parameter Gradient Estimator Because a discrete Bernoulli random sampling process is used to produce the configurations under which the log-likelihood of data samples is obtained, we require a gradient estimator to propagate gradient through to the $\\gamma$ structural meta-parameters. ",
+ "bbox": [
+ 174,
+ 839,
+ 825,
+ 882
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "We adopt for this purpose the gradient estimate of Bengio et al. (2019) (but see footnote 1): ",
+ "bbox": [
+ 171,
+ 103,
+ 756,
+ 118
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "img_path": "images/05663c44b32bacf4e94ac7b3c8468e7e7e7174553d7cf2a1c4855974bd17f26a.jpg",
+ "text": "$$\ng _ { i j } = \\frac { \\sum _ { k } ( \\sigma ( \\gamma _ { i j } ) - c _ { i j } ^ { ( k ) } ) \\mathcal { L } _ { C , i } ^ { ( k ) } ( X ) } { \\sum _ { k } \\mathcal { L } _ { C , i } ^ { ( k ) } ( X ) }\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 380,
+ 122,
+ 616,
+ 164
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where the $( k )$ superscript indicates the values obtained for the $k$ -th draw of $C$ . This gradient is estimated solely with $\\theta _ { \\mathrm { s l o w } }$ because estimates employing $\\theta$ have much greater variance. ",
+ "bbox": [
+ 176,
+ 170,
+ 821,
+ 199
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Acyclic Constraint We add a regularization term to the loss term that discourages the model from having length-2 cycles in the learned adjacency matrix. The regularizer term is ",
+ "bbox": [
+ 173,
+ 209,
+ 825,
+ 238
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "img_path": "images/c9e6c645bf54bf607c6162b9ac69233dcfc2d763372ff5553590f369bc6a63c0.jpg",
+ "text": "$$\nJ _ { \\mathrm { D A G } } = \\sum _ { i \\neq j } \\cosh ( \\sigma ( \\gamma _ { i j } ) \\sigma ( \\gamma _ { j i } ) ) , \\quad \\forall i , j \\in \\{ 0 , \\ldots , M - 1 \\}\n$$",
+ "text_format": "latex",
+ "bbox": [
+ 303,
+ 243,
+ 694,
+ 279
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "and is derived from Zheng et al. (2018). The details of the derivation are in the Appendix. ",
+ "bbox": [
+ 171,
+ 284,
+ 759,
+ 299
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.4 MODEL DESCRIPTION ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 315,
+ 369,
+ 329
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Learner Model We model the $M$ structural assignments $X _ { i } : = f _ { i } ( X _ { p a ( i , C ) } , N _ { i } )$ (Eq. 1) of the SCM as a set of $M$ multi-layer perceptrons (MLPs), as in Bengio et al. (2019). The MLPs are identical in shape but do not share any parameters, since they are modeling independent causal mechanisms. Each possesses an input layer of $M \\times N$ neurons (for $M$ one-hot vectors of length $N$ each), a single hidden layer chosen arbitrarily to have $\\operatorname* { m a x } ( 4 M , 4 N )$ neurons with a LeakyReLU activation of slope 0.1, and an output layer of $N$ neurons representing the unnormalized log-probabilities of each category. To force $f _ { i }$ to rely exclusively on the direct ancestor set $p a ( i , C )$ under adjacency matrix $C$ (See Eqn. 2), the one-hot input vector $X _ { j }$ for variable $X _ { i }$ ’s MLP is masked by the Boolean element $c _ { i j }$ . The functional parameters of the MLP are the set $\\theta = \\left\\{ { \\cal W } 0 _ { i h j n } , { \\tt B } 0 _ { i h } , { \\tt W } 1 _ { i n h } , { \\tt B } 1 _ { i n } \\right\\}$ . ",
+ "bbox": [
+ 173,
+ 340,
+ 826,
+ 467
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "An example of the multi-MLP architecture with $M = 3$ categorical variables of $N = 2$ categories is shown in Figure 1. ",
+ "bbox": [
+ 176,
+ 472,
+ 823,
+ 501
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Ground-Truth Model In our experiments, ground-truth SCM models are parametrized as a set of MLPs of the same size as the learner models, thus avoid having to manually define the Conditional Probability Tables (CPTs). Ground-truth models exist in two variants that differ mainly in initialization: synthetic, where the $\\theta$ are randomly initialized and the $\\gamma$ are either randomly-initialized or pre-specified; and real-world, where the $\\theta$ and $\\gamma$ are both initialized so as to closely replicate the CPTs of given Bayesian networks. ",
+ "bbox": [
+ 174,
+ 516,
+ 825,
+ 599
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.5 INTERVENTIONS ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 616,
+ 330,
+ 631
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Soft interventions To execute an intervention on variable $X _ { i }$ , we reinitialize $X _ { i }$ ’s ground-truth MLP parameters randomly while leaving other variables’ MLPs untouched. A copy of the old parameters is saved, allowing the intervention to be undone by resetting the parameters back to their original values. ",
+ "bbox": [
+ 174,
+ 642,
+ 825,
+ 699
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Predicting interventions After an intervention on $X _ { i }$ , the gradients into the learned model’s $\\gamma _ { i }$ and the slow weights for the $i$ -th conditional are false, because they do not bear the blame for $X _ { i }$ ’s outcome (which lies with the intervener). We find that ignoring this issue considerably hurts or slows down meta-learning, suggesting that we should try to infer on which variable the intervention took place. For this purpose, we take advantage of the fact that the conditional likelihood of the intervened variable tends to have a poorer relative likelihood under $D _ { \\mathrm { i n t } }$ , so we pick the variable with the greatest deterioration in likelihood as our guess. ",
+ "bbox": [
+ 174,
+ 713,
+ 825,
+ 811
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.6 TRAINING ALGORITHM ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 828,
+ 375,
+ 842
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "The structural meta-parameters $\\sigma ( \\gamma _ { i j } )$ represent the belief in the hypothesis that node $i$ has node $j$ as a direct causal parent. We may sample from this belief, obtaining different configurations (causal structures) of the causal graph. We hypothesize that the correct configuration enables better adaptation to a slight change in distribution, e.g. resulting from a soft intervention. Hence, we evaluate different configurations under the transfer distribution; those giving a higher transfer likelihood under $\\theta _ { \\mathrm { s l o w } }$ get a higher reward and their probability is increased. The functional (meta-)parameters are trained as usual by gradient ascent on the log-likelihood. The details of the training algorithm is in Section A.2 in the Appendix. ",
+ "bbox": [
+ 174,
+ 853,
+ 823,
+ 924
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 176,
+ 103,
+ 823,
+ 146
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Synthetic datasets use either a specified edge structure for $\\gamma$ , or randomly initialize $\\gamma _ { i j }$ such that it is Boolean, strictly lower-triangular, and each row has an expected sum of 1-5 (and therefore each node expects 1-5 direct ancestors). Real-world datasets, specified as CPTs, must first be converted or approximated by a near-ground-truth MLP. We use the graph’s proper edge structure to initialize $\\gamma$ and learn $\\theta$ by training individually each MLP in the set to replicate the correct pre-softmax logits. In practice, excellent reproductions of the CPT can be achieved. ",
+ "bbox": [
+ 174,
+ 152,
+ 825,
+ 236
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Stability of training Our model requires simultaneous training of both the structural and the functional meta-parameters, but these are not independent and do influence each other, which leads to instability in training. For example, if $\\sigma ( \\gamma _ { i j } ) \\approx 0$ incorrectly, the $i$ -th MLP does not learn to use input $X _ { j }$ , and vice-versa, if the $i$ -th MLP has not learned to properly use input $X _ { j }$ , this will favour pushing $\\sigma ( \\gamma _ { i j } )$ towards 0. To overcome this instability, we pretrain the model under observational data (from the distribution of the data before interventions) using dropout on the inputs. This ensures that the functional meta-parameters $\\theta _ { \\mathrm { s l o w } }$ are not too biased towards certain configurations of the meta-parameters $\\gamma$ . ",
+ "bbox": [
+ 174,
+ 243,
+ 825,
+ 354
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4 RELATED WORK ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 366,
+ 344,
+ 382
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "The recovery of the underlying structural causal graph from observational and interventional data is a fundamental problem (Pearl, 1995; 2009). Different approaches have been studied, score-based, constraint-based and asymmetry-based methods. Score-based methods search through the space of all possible directed acyclic graphs (DAGs) representing the causal structure based on some form of scoring function for network structures (Chickering, 2002; Tsamardinos et al., 2006; Goudet et al., 2017; Hauser & Bühlmann, 2012; Heckerman et al., 1995; Cooper & Yoo, 1999; Hauser & Bühlmann, 2012). Constraint-based methods (Spirtes et al., 2000; Sun et al., 2007; Zhang et al., 2012) infer the DAG by analyzing the conditional independence of data. Eaton & Murphy (2007b) use dynamic programming techniques to accelerate Markov Chain Monte Carlo (MCMC) sampling in a Bayesian approach to structure learning for discrete variable DAGs. Asymmetry-based methods (Shimizu et al., 2006; Hoyer et al., 2009; Daniusis et al., 2012; Budhathoki & Vreeken, 2017; Mitrovic et al., 2018) assume asymmetry between cause and effect in the data and try to use this information to estimate the causal structure. Recently Peters et al. (2016); Ghassami et al. (2017) proposed to exploit invariance across different environments to infer causal structure, but are difficult to scale to large graphs due to the necessary iteration over the super-exponential set of possible graphs. ",
+ "bbox": [
+ 174,
+ 388,
+ 826,
+ 597
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "For interventional data, it is often assumed that the models have access to full intervention information, which is rare in the real world. Rothenhäusler et al. (2015) have investigated the case of additive shift interventions, while Eaton & Murphy (2007a) have examined the situation where the targets of experimental interventions are imperfect or uncertain. This is different from our setting where the intervention is unknown to start with and is assumed to arise from other agents and the environment. ",
+ "bbox": [
+ 174,
+ 603,
+ 825,
+ 672
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Learning based methods have been proposed (Guyon, a;b; Lopez-Paz et al., 2015) and there also exist recent approaches using the generalization ability of neural networks to learn causal signals from purely observational data (Kalainathan et al., 2018; Goudet et al., 2018). Neural network methods equipped with learned masks, such as (Ivanov et al., 2018; Li et al., 2019; Yoon et al., 2018; Douglas et al., 2017), exist in the literature, but only a few (Kalainathan et al., 2018) have been adapted to causal inference. This last work is, however, tailored for causal inference on continuous variables and from observations only. Adapting it to a discrete-variable setting is made difficult by its use of a Generative Adversarial Network (GAN) Goodfellow et al. (2014) framework. ",
+ "bbox": [
+ 174,
+ 680,
+ 825,
+ 791
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Turning now to meta-learning, Dasgupta et al. (2019) have used it to learn to make predictions under interventions. However, their approach does not induce a causal graph, neither explicitly nor via decoding. Thus, it cannot be used for general causal discovery, but only to make predictions of variable values. Most similar to our work, Bengio et al. (2019) proposes a meta-learning framework for learning causal models from interventional data. However, the proposed method (Bengio et al., 2019) explicitly models every possible set of parents for every child variable and attempts to distinguish the best among them. Because there are combinatorially-many such parent sets, the method cannot scale beyond trivial graphs. In our work, we bypass this restriction by modeling the edge between any 2 variables as a dropout probability and hence our model only scales quadratically with the graph size. ",
+ "bbox": [
+ 174,
+ 797,
+ 825,
+ 924
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "5 EXPERIMENTAL SETUP AND RESULTS ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 103,
+ 516,
+ 117
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Our experiments aim to evaluate the proposed method to recover the correct causal structure and which elements of the method matter. We first evaluate our model on a synthetic dataset where we have control over the number of variables and causal edges in the ground-truth SCM. This allows us to verify after learning with what accuracy we recover the individual decisions about $c _ { i j }$ and understand the performance of our algorithm under various conditions. We then evaluate our method on real world datasets collected from the BnLearn dataset repository, and show that the proposed approach recovers the true causal structure. We then show that the trained models can correctly predict the consequences of previously unseen interventions on the rest of the graph. We also perform ablations showing how important is each component of the model. ",
+ "bbox": [
+ 173,
+ 126,
+ 825,
+ 251
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "5.1 SYNTHETIC DATASETS ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 267,
+ 372,
+ 280
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/f99f187fc376946f89ba5b065c406dc52459b277e1a3d71f66784cd8ed754b6a.jpg",
+ "image_caption": [
+ "Figure 2: Learned edges at three different stages of training. Left: Chain graph with 4 variables. Right: Fully-connected DAG graph with 4 variables. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 263,
+ 284,
+ 730,
+ 337
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "We first evaluate the model’s performance on several randomly-initialized SCMs with specific, representative graph structures. For $M = 3$ -variable DAGs, we consider every possible connected graph: chain3, fork3, collider3 and confounder3 (See Fig. 7 in appendix). They exhibit every graph sub-structure that can exist in larger graphs, and must be mastered before tackling larger graphs. Since the number of possible DAGs grows super-exponentially with the number of variables, for $M > 3$ up to 8 a selection of representative and edge-case graphs are chosen. The chainM and fullM graphs $( \\mathbb { M } = 3 - 8$ ) are the minimally and maximally connected M-variable graphs, while the remaining graphs are randomly generated with a varying sparsity level (1-4 expected number of parents per node). The details of the setup can be found in Appendix A.5. ",
+ "bbox": [
+ 173,
+ 373,
+ 825,
+ 498
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Results The model can successfully recover the correct edge structure for all synthetic graphs considered. The learning curves plotting the average cross-entropy (CE) loss for the learned edges against the ground-truth model for $M = 3$ are shown in Figure 7. The fully-connected confounder3 graph is particularly easy to learn, but all 3-variable graphs are learned perfectly. Plots of the same process on 4- through 8-variable graphs were similarly encouraging, with all models converging to a negligible loss. The results are, however, sensitive to some hyperparameters, notably the DAG penalty and the sparsity penalty. ",
+ "bbox": [
+ 173,
+ 512,
+ 825,
+ 611
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "5.2 REAL-WORLD DATASETS: BNLEARN ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 626,
+ 472,
+ 640
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "The Bayesian Network Repository is a collection of commonlyused causal Bayesian networks from the literature, suitable for Bayesian and causal learning benchmarks. We evaluate our model on the Earthquake (Korb & Nicholson, 2010), Cancer (Korb & Nicholson, 2010) and Asia (Lauritzen & Spiegelhalter, 1988) datasets ( $M = 5$ , 5 and 8-variables respectively, maximum 2 parents per node) in the BnLearn dataset repository. ",
+ "bbox": [
+ 174,
+ 652,
+ 583,
+ 750
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "The ground-truth SCM is given for each dataset, and the functional parameters are represented as conditional probability tables (CPTs). We learn a near-ground-truth MLP from the dataset’s CPT and use it as the ground-truth data generator. We also insert a (greater than 1) temperature factor in order to increase the likelihood of sampling some very rare events in the CPTs. Details of the setup can be found in Appendix A.5.1. ",
+ "bbox": [
+ 174,
+ 751,
+ 825,
+ 819
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/d9b5c523ea9b4196e4b0c1585be866e7e439433e299ccc6ecc2da33b588016ef.jpg",
+ "image_caption": [
+ "Figure 3: Earthquake: Learned edges at three different stages of training. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 596,
+ 656,
+ 821,
+ 713
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Results The model can successfully recover the correct edge structure for all BnLearn graphs considered up to and including 8-variable Asia. Figures 3 and 4 illustrate what the model has learned at several stages of learning. In these figures, the $\\sigma ( \\gamma _ { i j } )$ and $c _ { i j }$ adjacency matrix elements are plotted as colored squares and dots. Off-diagonal terms are unknown, and appear yellow. ",
+ "bbox": [
+ 174,
+ 827,
+ 581,
+ 922
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/1f0d4465e558dd224c900b27c6e12284050ff546dc987c0d1ce52780e5de9750.jpg",
+ "image_caption": [
+ "Figure 4: Asia: Learned edges at three different stages of training. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 598,
+ 829,
+ 820,
+ 886
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/4233c6c69310a6d99481788ea1053d2c1bf7319a29ae8964709cdf0d3f64ee2c.jpg",
+ "image_caption": [
+ "Figure 5: Left: Cross entropy (CE) for edge probability between learned and ground-truth graphs for all 3-variable SCMs. Error bars are $\\pm 1 \\sigma$ over PRNG seeds 1-5. Middle: Edge CE loss for the chain graph with 4-7 variables. Right: Edge CE loss for 3-variable graphs with no dropout during pretraining, showing the importance of this dropout. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 210,
+ 89,
+ 789,
+ 202
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/62682ff61ea32624cedad428b27f87926c737a7dfde75025a4c006be7dbb5857.jpg",
+ "image_caption": [
+ "Figure 6: Ablations study results on all possible 3 variable graphs. Both graphs show the cross-entropy loss on learned vs ground-truth edges over training time. Left: Models that infer the intervention (prediction, bold) vs models that have knowledge of the true intervention (ground truth, long dash) vs models that use no knowledge of the intervention at all (no prediction, short dash). Result suggests inferring the intervention works almost as well as knowing the true intervention. Right: Comparisons of model trained with and without DAG regularizer $( L _ { \\mathrm { D A G } } )$ , showing that DAG regularizer helps convergence. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 307,
+ 258,
+ 689,
+ 369
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "The color of the squares indicates belief in the presence or \nabsence of an edge $( \\sigma ( \\gamma _ { i j } ) )$ , while the color of the dot indicates the ground truth $c _ { i j }$ . Red indicates (belief in) an edge; Blue indicates (belief in) the absence of an edge. Yellow indicates maximum uncertainty At the beginning of training, the main diagonal is a priori known to be clear of edges, and therefore is solid blue. ",
+ "bbox": [
+ 174,
+ 453,
+ 825,
+ 522
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "As training progresses, the beliefs approach the ground truth, which visually appears as the squares converging towards the color of the dot within them. When they coincide, the dot vanishes. An erroneous belief stands out as a red dot on blue square or, vice-versa, a blue dot on red square. Because we pre-sort the nodes in BnLearn datasets so that they are in topological order, the model must learn a lower triangle. This corresponds to a completely blue uppper triangle. Anti-causal violations are easily recognizable as non-blue squares in the upper triangle. ",
+ "bbox": [
+ 174,
+ 530,
+ 825,
+ 613
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Baseline comparisons We compared our method to ICP (Peters et al., 2016) and Eaton & Murphy (2007a). Eaton & Murphy (2007a) handles uncertain interventions and Peters et al. (2016) handles unknown interventions. However, neither attempt to predict the intervention. ",
+ "bbox": [
+ 176,
+ 630,
+ 820,
+ 672
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "table",
+ "img_path": "images/4c32a7f801ee7bdc5f66b7476763d940a601edf7f750c0cb91ff1a5ec29e6300.jpg",
+ "table_caption": [
+ "Table 1: Baseline comparisons: Cross entropy (lower is better) for edge probability on learned and ground-truth edges on Asia graph. compared to to Peters et al. (2016), (Eaton & Murphy, 2007a) and (Zheng et al., 2018) "
+ ],
+ "table_footnote": [],
+ "table_body": "| Algorithm1Training Algorithm |
| 1: procedure TRAINING(Categorical Distribution D,with M nodes and N categories) Let ian integer from O to M-1 |
| 2: 34 | |
| for kpretrain steps do |
| 5: | x~D |
| 6: | c ~ Ber(σ(γ)) |
| 7: 8: | L = -log P(xlc) > Compute log-probability of data given config |
| 0slow ←Adam(0slow,VL) |
| 9: for kintervention steps do | |
| 10: | I_N←randint(O，M-1) |
| 11: | Dint := D with intervention on node I_N |
| 12: | if predicting intervention then |
| 13: | Li←O∀i |
| 14: | for Kpredict steps do |
| 15: | x~Dint |
| 16: | c ~ Ber(σ(γ)) |
| 17: 18: | Li←Li+-logPi(xlci;Oslow)∀i |
| >Accumulate NLL for every node i separately I_N←argmax(Li) |
| |
| 19: 20: | gammagrads，logregrets =[]，[] >Transfer Episode Adaptation Loop |
| 21: | for Kepisode steps do x~Dint |
| 22: | gammagrad，logregret = 0，0 |
| 23: | for kcfg steps do |
| 24: | c ~ Ber(σ(γ)) |
| 25: | Li=-log Pi(xlci;0slow）∀i |
| 26: | gammagrad += σ(γ)-c |
| 27: | logregret += ∑ Li |
| 28: | I_N |
| 29: | gammagrads.append(gammagrad) logregrets.append(logregret) |
| |
| 30: | J ←入MaxEnt LMaxEnt(γ)+ 入sparse LSparse(γ)+ 入DAG LDAG(γ) |
| 31: | Vγ←VγJ+∑gammagradskijlogregrets.softmax(0) ki |
| 32: | k γ ← Adam(γ,∀γ) |
",
+ "bbox": [
+ 171,
+ 156,
+ 826,
+ 661
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "A.3 PRELIMINARIES ",
+ "text_level": 1,
+ "bbox": [
+ 174,
+ 688,
+ 330,
+ 702
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Interventions In a purely-observational setting, it is known that causal graphs can be distinguished only up to a Markov equivalence class. In order to identify the true causal graph intervention data is needed (Eberhardt et al., 2012). Several types of common interventions may be available (Eaton & Murphy, 2007a). These are: No intervention: only observational data is obtained from the ground truth causal model. Hard/perfect: the value of a single or several variables is fixed and then ancestral sampling is performed on the other variables. Soft/imperfect: the conditional distribution of the variable on which the intervention is performed is changed. Uncertain: the learner is not sure of which variable exactly the intervention affected directly. Here we make use of soft interventions for several reasons: First, they include hard interventions as a limiting case and hence are more general. Second, in many real-world scenarios, it is more difficult to perform a hard intervention compared to a soft one. We also deal with a special case of uncertain interventions, where the variable selected for intervention is random and unknown. We call these unidentified or unknown interventions. ",
+ "bbox": [
+ 173,
+ 707,
+ 825,
+ 873
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Causal sufficiency The inability to distinguish which causal graph, within a Markov equivalence class, is the correct one in the purely-observational setting is called the identifiability problem. In our setting, all variables are observed (there are no latent confounders) and all interventions are random and independent. Hence, within our setting the true causal graph is always identifiable in principle (Eberhardt et al., 2012; Heinze-Deml et al., 2018). We consider here situations where a single variable is randomly selected and intervened upon with a soft or imprecise intervention, its identity is unknown and must be inferred. ",
+ "bbox": [
+ 176,
+ 882,
+ 823,
+ 924
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "",
+ "bbox": [
+ 174,
+ 103,
+ 825,
+ 159
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "A.4 EXPERIMENTAL SETUP ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 181,
+ 377,
+ 195
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "For all datasets, the weight parameters for the learned model is initialized randomly. In order to not bias the structural parameters, all $\\gamma$ is initialized to 0.5 in the beginning of training. ",
+ "bbox": [
+ 174,
+ 208,
+ 823,
+ 237
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "A.5 SYNTHETIC DATA ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 257,
+ 339,
+ 272
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "SCM with $n$ variables is modeled by $n$ feedforward neural networks (MLPs) as described in section 3.1. For simplicity, we assume use an acyclic causal graph such that we could easily sample from it. Hence, given any pair of random variables $A$ and $B$ , either $A B$ , $B A$ or $A$ and $B$ are independent. ",
+ "bbox": [
+ 174,
+ 285,
+ 825,
+ 340
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "The MLP representing the ground-truth SCM has its weights $\\theta$ initialized use orthogonal initialization with gain 2.5 and the biases are initialized using a uniform initialization between $- 1 . 1$ and 1.1, which was empirically found to yield \"interesting\" yet learnable random SCMs. ",
+ "bbox": [
+ 174,
+ 348,
+ 825,
+ 390
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "image",
+ "img_path": "images/3a790701f5c12dec0fe6dde4d2cd704750cc5a7f50499f417a7dd1e4510253c9.jpg",
+ "image_caption": [
+ "Figure 7: Left: Every possible 3-variable connected DAG. Right: Cross entropy for edge probability between learned and ground-truth SCM for all 3-variable SCMs. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 184,
+ 421,
+ 751,
+ 540
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "A.5.1 BNLEARN DATA REPOSITORY ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 608,
+ 433,
+ 622
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "The repo contains many datasets with various sizes and structures modeling different variables. We evaluate our model on 3 of the datasets in the repo, namely the Earthquake (Korb & Nicholson, 2010), Cancer (Korb & Nicholson, 2010) and Asia (Lauritzen & Spiegelhalter, 1988) datasets. The ground-truth model structure for the Cancer (Korb & Nicholson, 2010) and Earthquake (Korb & Nicholson, 2010) datasets are shown in Figure 8. Note that even though the structure for the 2 datasets seems to be the same, the conditional probability tables (CPTs) for these 2 datasets are very different and hence results in different structured causal models (SCMs) for the 2 datasets. ",
+ "bbox": [
+ 173,
+ 632,
+ 826,
+ 731
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "image",
+ "img_path": "images/74b53bb5cabd72d9b02acd007967754ee13b35decbfe1cd63be56773fcce91d8.jpg",
+ "image_caption": [
+ "Figure 8: Left: Ground Truth SCM for Cancer. Middle: Groundtruth SCM for Earthquake. Right: Groundtruth SCM for Asia. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 261,
+ 758,
+ 736,
+ 897
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "A.5.2 TRAINING GROUND-TRUTH DATA GENERATOR ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 103,
+ 545,
+ 117
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Because a CPT is capable of representing any distribution, and MLPs are strictly less powerful in this respect, it may not be possible to learn perfectly the distribution with our MLP learner model. We therefore train a near-ground-truth MLP to replicate as closely as possible the CPT’s probability table, and then use this trained MLP are the ground-truth SCM data generator. ",
+ "bbox": [
+ 174,
+ 127,
+ 825,
+ 184
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Training is by 1000 iterations of full-batch gradient descent with learning rate 0.001 and momentum 0.9, with all possible parent values masked with the ground-truth vector $\\gamma _ { i }$ . The objective is to minimize mean squared error between the MLP’s logits and the log-probability as drawn from the CPT. If the CPT contains a zero, it is approximated by a logit of $- 1 0 0$ . ",
+ "bbox": [
+ 173,
+ 190,
+ 825,
+ 247
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Given that some of the CPTs contain very unlikely events, we have found it necessary to add a temperature parameter in order to make them more frequent. The near-ground-truth MLP model’s logit outputs are divided by the temperature before being used for sampling. Temperatures above 1 result in more uniform distributions for all causal variables; Temperatures below 1 result in less uniform, sharper distributions that peak around the most likely value. We find empirically that a temperature of about 2 is required for our BnLearn benchmarks. ",
+ "bbox": [
+ 174,
+ 253,
+ 825,
+ 337
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "A.6 EFFECT OF SPARSITY ",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 353,
+ 366,
+ 367
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "We use a $L 1$ regularizer on the structure parameters $\\gamma$ to encourage a sparse representation of edges in the causal graph. In order to better understand the effect of the $L 1$ regularizer, we conducted ablation studies on the $L 1$ regularizer. It seems that the regularizer has an small effect on rate of converges and that the model converges faster with the regularizer. This is shown in Figure 9 ",
+ "bbox": [
+ 174,
+ 381,
+ 825,
+ 435
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "image",
+ "img_path": "images/db53278365d2e1de3d89b57ce0a66803a6f2763f2c38f0b3d98a6e5b040f87b2.jpg",
+ "image_caption": [
+ "Figure 9: Effect of Sparsity: On 5 variable, 6 variable and 8 variable Nodes "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 202,
+ 458,
+ 784,
+ 563
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "A.7 EFFECT OF TEMPERATURE",
+ "text_level": 1,
+ "bbox": [
+ 176,
+ 606,
+ 401,
+ 619
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "As noted in section A.5.1, we have introduced a temperature hyperparameter in order to encourage the groundtruth model to generate some very rare events in the conditional probability tables (CPTs) more frequently. We run ablation studies to understand the importance of the temperature term. A temperature of 1 corresponds to no changes to the underlying CPTs. As shown in Figure 12, for the Cancer (Korb & Nicholson, 2010) dataset, a temperature of 2 improves the accuracy of causal graph recovery. ",
+ "bbox": [
+ 173,
+ 632,
+ 825,
+ 715
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "image",
+ "img_path": "images/d8967b64efd9a61747efc4636ecfe6584f61e715cdd46ad48ac565ed8ddfd481.jpg",
+ "image_caption": [
+ "Figure 10: Cross entropy for edge probability between learned and ground-truth SCM for Cancer at varying temperatures. "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 367,
+ 743,
+ 614,
+ 886
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "image",
+ "img_path": "images/93a87e2e595b8857314a6cccac873b3eb129f202923fb425b9075b306cf5a5f5.jpg",
+ "image_caption": [
+ "Figure 11: Cross entropy for edge probability between learned and ground-truth SCM. Left: The Earthquake dataset with 6 variables. Right: The Asia dataset with 8 variables "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 227,
+ 213,
+ 763,
+ 363
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "image",
+ "img_path": "images/97ce72bf68cf2b4152045f11fbfda91a63a4d62a35f39a1aa1fc0b78630d6cf9.jpg",
+ "image_caption": [
+ "Figure 12: Left: SCM for cross5 graph. Right: Cross entropy for edge probability between learned and ground-truth SCM for chain5 and cross5, comparing the learning process with prediction of edges to ground-truth intervention "
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 245,
+ 606,
+ 766,
+ 787
+ ],
+ "page_idx": 16
+ }
+]
\ No newline at end of file
diff --git a/parse/train/H1gN6kSFwS/H1gN6kSFwS_middle.json b/parse/train/H1gN6kSFwS/H1gN6kSFwS_middle.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba23b8ad90b2157d01301570cad6765a0477d417
--- /dev/null
+++ b/parse/train/H1gN6kSFwS/H1gN6kSFwS_middle.json
@@ -0,0 +1,45975 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 141,
+ 78,
+ 470,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 77,
+ 472,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 77,
+ 472,
+ 97
+ ],
+ "score": 1.0,
+ "content": "LEARNING NEURAL CAUSAL MODELS FROM",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 202,
+ 98,
+ 410,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 202,
+ 98,
+ 410,
+ 117
+ ],
+ "score": 1.0,
+ "content": "UNKNOWN INTERVENTIONS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 135,
+ 245,
+ 157
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 133,
+ 201,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 133,
+ 201,
+ 149
+ ],
+ "score": 1.0,
+ "content": "Anonymous authors",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 111,
+ 146,
+ 245,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 146,
+ 245,
+ 158
+ ],
+ "score": 1.0,
+ "content": "Paper under double-blind review",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 178,
+ 333,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 177,
+ 335,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 177,
+ 335,
+ 191
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 204,
+ 469,
+ 423
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 205,
+ 469,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 205,
+ 469,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Meta-learning over a set of distributions can be interpreted as learning different",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 215,
+ 470,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 215,
+ 470,
+ 227
+ ],
+ "score": 1.0,
+ "content": "types of parameters corresponding to short-term vs long-term aspects of the mech-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 226,
+ 469,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 226,
+ 469,
+ 239
+ ],
+ "score": 1.0,
+ "content": "anisms underlying the generation of data. These are respectively captured by",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 237,
+ 470,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 237,
+ 470,
+ 250
+ ],
+ "score": 1.0,
+ "content": "quickly-changing parameters and slowly-changing meta-parameters. We present a",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 248,
+ 469,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 248,
+ 469,
+ 260
+ ],
+ "score": 1.0,
+ "content": "new framework for meta-learning causal models where the relationship between",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 259,
+ 470,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 259,
+ 470,
+ 271
+ ],
+ "score": 1.0,
+ "content": "each variable and its parents is modeled by a neural network, modulated by struc-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 142,
+ 270,
+ 469,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 270,
+ 469,
+ 282
+ ],
+ "score": 1.0,
+ "content": "tural meta-parameters which capture the overall topology of a directed graphical",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 142,
+ 281,
+ 470,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 281,
+ 470,
+ 293
+ ],
+ "score": 1.0,
+ "content": "model. Our approach avoids a discrete search over models in favour of a continu-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 292,
+ 469,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 292,
+ 469,
+ 303
+ ],
+ "score": 1.0,
+ "content": "ous optimization procedure. We study a setting where interventional distributions",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 303,
+ 470,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 303,
+ 470,
+ 315
+ ],
+ "score": 1.0,
+ "content": "are induced as a result of a random intervention on a single unknown variable of",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 314,
+ 469,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 314,
+ 469,
+ 326
+ ],
+ "score": 1.0,
+ "content": "an unknown ground truth causal model, and the observations arising after such",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 140,
+ 323,
+ 470,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 323,
+ 470,
+ 339
+ ],
+ "score": 1.0,
+ "content": "an intervention constitute one meta-example. To disentangle the slow-changing",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 336,
+ 470,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 336,
+ 470,
+ 348
+ ],
+ "score": 1.0,
+ "content": "aspects of each conditional from the fast-changing adaptations to each intervention,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 142,
+ 348,
+ 470,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 348,
+ 470,
+ 358
+ ],
+ "score": 1.0,
+ "content": "we parametrize the neural network into fast parameters and slow meta-parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 357,
+ 470,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 357,
+ 470,
+ 370
+ ],
+ "score": 1.0,
+ "content": "We introduce a meta-learning objective that favours solutions robust to frequent but",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 369,
+ 469,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 369,
+ 469,
+ 381
+ ],
+ "score": 1.0,
+ "content": "sparse interventional distribution change, and which generalize well to previously",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 380,
+ 469,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 380,
+ 469,
+ 392
+ ],
+ "score": 1.0,
+ "content": "unseen interventions. Optimizing this objective is shown experimentally to recover",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 390,
+ 469,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 390,
+ 469,
+ 402
+ ],
+ "score": 1.0,
+ "content": "the structure of the causal graph. Finally, we find that when the learner is unaware",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 401,
+ 469,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 401,
+ 469,
+ 414
+ ],
+ "score": 1.0,
+ "content": "of the intervention variable, it is able to infer that information, improving results",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 412,
+ 457,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 412,
+ 457,
+ 425
+ ],
+ "score": 1.0,
+ "content": "further and focusing the parameter and meta-parameter updates where needed.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 448,
+ 206,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 208,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 208,
+ 463
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 473,
+ 505,
+ 594
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 506,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 506,
+ 487
+ ],
+ "score": 1.0,
+ "content": "A major challenge of contemporary deep learning is to generalize well outside the assumptions of",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 485,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 485,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "independent and identically distributed data, when we care about generalization or fast adaptation",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "to distributions other than the main training distribution. For this purpose, we propose an approach",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 506,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 506,
+ 520
+ ],
+ "score": 1.0,
+ "content": "that starts by distinguishing between: (a) an underlying causal model, (b) observational distributions",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 517,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "derived from it, an (c) interventional distributions arising from interventions upon its variables,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "whether they be known or unknown. Fortunately, these distinctions can be addressed by the paradigm",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "of Structural Causal Models (SCMs) (Pearl, 1995; Peters et al., 2017) and a wide body of associated",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "literature. Unlike other frameworks using SCMs we also account for interventions performed",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "by agents other than an experimenter. By treating the interventions of other agents as unknown",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "interventions that lead to changes in the underlying data distribution, the present work is a contribution",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 582,
+ 357,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 357,
+ 595
+ ],
+ "score": 1.0,
+ "content": "towards the use of a meta-learning for causal model induction.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 600,
+ 506,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "Estimating the underlying causal structure from data is an open and challenging problem (Pearl,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "2009; Imbens & Rubin, 2015). A lot of prior work has examined learning causal structure based",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "on observational data (Chickering, 2002; Tsamardinos et al., 2006; Goudet et al., 2017; Hauser &",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 507,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 507,
+ 646
+ ],
+ "score": 1.0,
+ "content": "Bühlmann, 2012; Spirtes et al., 2000; Sun et al., 2007; Zhang et al., 2012; Shimizu et al., 2006;",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "Hoyer et al., 2009; Daniusis et al., 2012; Budhathoki & Vreeken, 2017; Kalainathan et al., 2018).",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "However, many real-world datasets have an inherent distributional heterogeneity due to different",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "interventions to the variables composing the model. In these situations interventional approaches are",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "needed (e.g. Heckerman et al., 1995; Cooper & Yoo, 1999; Hauser & Bühlmann, 2012; Peters et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "2016; Rothenhäusler et al., 2015; Ghassami et al., 2017). Established approaches for causal inference",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "are often either relying on restrictive assumptions or on conditional independence testing, which is",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "hard (Shah & Peters, 2018). Furthermore, most of these approaches either assume full knowledge of",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 456,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 456,
+ 733
+ ],
+ "score": 1.0,
+ "content": "the intervention or make strong assumptions about its form (Heinze-Deml et al., 2018).",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 761
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 141,
+ 78,
+ 470,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 77,
+ 472,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 77,
+ 472,
+ 97
+ ],
+ "score": 1.0,
+ "content": "LEARNING NEURAL CAUSAL MODELS FROM",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 202,
+ 98,
+ 410,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 202,
+ 98,
+ 410,
+ 117
+ ],
+ "score": 1.0,
+ "content": "UNKNOWN INTERVENTIONS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 135,
+ 245,
+ 157
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 133,
+ 201,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 133,
+ 201,
+ 149
+ ],
+ "score": 1.0,
+ "content": "Anonymous authors",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 111,
+ 146,
+ 245,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 146,
+ 245,
+ 158
+ ],
+ "score": 1.0,
+ "content": "Paper under double-blind review",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 111,
+ 133,
+ 245,
+ 158
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 178,
+ 333,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 177,
+ 335,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 177,
+ 335,
+ 191
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 204,
+ 469,
+ 423
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 205,
+ 469,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 205,
+ 469,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Meta-learning over a set of distributions can be interpreted as learning different",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 215,
+ 470,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 215,
+ 470,
+ 227
+ ],
+ "score": 1.0,
+ "content": "types of parameters corresponding to short-term vs long-term aspects of the mech-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 226,
+ 469,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 226,
+ 469,
+ 239
+ ],
+ "score": 1.0,
+ "content": "anisms underlying the generation of data. These are respectively captured by",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 237,
+ 470,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 237,
+ 470,
+ 250
+ ],
+ "score": 1.0,
+ "content": "quickly-changing parameters and slowly-changing meta-parameters. We present a",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 248,
+ 469,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 248,
+ 469,
+ 260
+ ],
+ "score": 1.0,
+ "content": "new framework for meta-learning causal models where the relationship between",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 259,
+ 470,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 259,
+ 470,
+ 271
+ ],
+ "score": 1.0,
+ "content": "each variable and its parents is modeled by a neural network, modulated by struc-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 142,
+ 270,
+ 469,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 270,
+ 469,
+ 282
+ ],
+ "score": 1.0,
+ "content": "tural meta-parameters which capture the overall topology of a directed graphical",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 142,
+ 281,
+ 470,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 281,
+ 470,
+ 293
+ ],
+ "score": 1.0,
+ "content": "model. Our approach avoids a discrete search over models in favour of a continu-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 292,
+ 469,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 292,
+ 469,
+ 303
+ ],
+ "score": 1.0,
+ "content": "ous optimization procedure. We study a setting where interventional distributions",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 303,
+ 470,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 303,
+ 470,
+ 315
+ ],
+ "score": 1.0,
+ "content": "are induced as a result of a random intervention on a single unknown variable of",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 314,
+ 469,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 314,
+ 469,
+ 326
+ ],
+ "score": 1.0,
+ "content": "an unknown ground truth causal model, and the observations arising after such",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 140,
+ 323,
+ 470,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 323,
+ 470,
+ 339
+ ],
+ "score": 1.0,
+ "content": "an intervention constitute one meta-example. To disentangle the slow-changing",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 336,
+ 470,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 336,
+ 470,
+ 348
+ ],
+ "score": 1.0,
+ "content": "aspects of each conditional from the fast-changing adaptations to each intervention,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 142,
+ 348,
+ 470,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 348,
+ 470,
+ 358
+ ],
+ "score": 1.0,
+ "content": "we parametrize the neural network into fast parameters and slow meta-parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 357,
+ 470,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 357,
+ 470,
+ 370
+ ],
+ "score": 1.0,
+ "content": "We introduce a meta-learning objective that favours solutions robust to frequent but",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 369,
+ 469,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 369,
+ 469,
+ 381
+ ],
+ "score": 1.0,
+ "content": "sparse interventional distribution change, and which generalize well to previously",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 380,
+ 469,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 380,
+ 469,
+ 392
+ ],
+ "score": 1.0,
+ "content": "unseen interventions. Optimizing this objective is shown experimentally to recover",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 390,
+ 469,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 390,
+ 469,
+ 402
+ ],
+ "score": 1.0,
+ "content": "the structure of the causal graph. Finally, we find that when the learner is unaware",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 401,
+ 469,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 401,
+ 469,
+ 414
+ ],
+ "score": 1.0,
+ "content": "of the intervention variable, it is able to infer that information, improving results",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 412,
+ 457,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 412,
+ 457,
+ 425
+ ],
+ "score": 1.0,
+ "content": "further and focusing the parameter and meta-parameter updates where needed.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 140,
+ 205,
+ 470,
+ 425
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 448,
+ 206,
+ 460
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 208,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 208,
+ 463
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 473,
+ 505,
+ 594
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 506,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 473,
+ 506,
+ 487
+ ],
+ "score": 1.0,
+ "content": "A major challenge of contemporary deep learning is to generalize well outside the assumptions of",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 485,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 485,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "independent and identically distributed data, when we care about generalization or fast adaptation",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "to distributions other than the main training distribution. For this purpose, we propose an approach",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 506,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 506,
+ 520
+ ],
+ "score": 1.0,
+ "content": "that starts by distinguishing between: (a) an underlying causal model, (b) observational distributions",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 517,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "derived from it, an (c) interventional distributions arising from interventions upon its variables,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "whether they be known or unknown. Fortunately, these distinctions can be addressed by the paradigm",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "of Structural Causal Models (SCMs) (Pearl, 1995; Peters et al., 2017) and a wide body of associated",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "literature. Unlike other frameworks using SCMs we also account for interventions performed",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "by agents other than an experimenter. By treating the interventions of other agents as unknown",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "interventions that lead to changes in the underlying data distribution, the present work is a contribution",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 582,
+ 357,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 357,
+ 595
+ ],
+ "score": 1.0,
+ "content": "towards the use of a meta-learning for causal model induction.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 105,
+ 473,
+ 506,
+ 595
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 600,
+ 506,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "Estimating the underlying causal structure from data is an open and challenging problem (Pearl,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "2009; Imbens & Rubin, 2015). A lot of prior work has examined learning causal structure based",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "on observational data (Chickering, 2002; Tsamardinos et al., 2006; Goudet et al., 2017; Hauser &",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 507,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 507,
+ 646
+ ],
+ "score": 1.0,
+ "content": "Bühlmann, 2012; Spirtes et al., 2000; Sun et al., 2007; Zhang et al., 2012; Shimizu et al., 2006;",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "Hoyer et al., 2009; Daniusis et al., 2012; Budhathoki & Vreeken, 2017; Kalainathan et al., 2018).",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "However, many real-world datasets have an inherent distributional heterogeneity due to different",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "interventions to the variables composing the model. In these situations interventional approaches are",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "needed (e.g. Heckerman et al., 1995; Cooper & Yoo, 1999; Hauser & Bühlmann, 2012; Peters et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "2016; Rothenhäusler et al., 2015; Ghassami et al., 2017). Established approaches for causal inference",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "are often either relying on restrictive assumptions or on conditional independence testing, which is",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "hard (Shah & Peters, 2018). Furthermore, most of these approaches either assume full knowledge of",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 456,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 456,
+ 733
+ ],
+ "score": 1.0,
+ "content": "the intervention or make strong assumptions about its form (Heinze-Deml et al., 2018).",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 105,
+ 600,
+ 507,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 181
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "However, in the real world, interventions are also not always performed by an experimenter. They",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "can be performed by other agents, or by environmental changes in ways that are unknown, or by",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "a naive learner (like a robot) which does not know precisely yet how its low-level actions change",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "high-level causal variables. In this paper, we look at the setting where interventions are unknown, and",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 126,
+ 507,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 126,
+ 507,
+ 140
+ ],
+ "score": 1.0,
+ "content": "our goal is to discover causal graphs given unknown-intervention samples. The challenging aspect of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 507,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 507,
+ 150
+ ],
+ "score": 1.0,
+ "content": "this setting is to not only learn the causal graph structure, but also predict the intervention accurately.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 147,
+ 506,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 506,
+ 161
+ ],
+ "score": 1.0,
+ "content": "In this setting, we need to make sure to: (1) avoid an exponential search over all possible DAGs, (2)",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 156,
+ 506,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 156,
+ 506,
+ 173
+ ],
+ "score": 1.0,
+ "content": "handle unknown interventions, (3) model the effect of interventions, and (4) model the underlying",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 170,
+ 173,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 170,
+ 173,
+ 182
+ ],
+ "score": 1.0,
+ "content": "causal structure.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 186,
+ 505,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 200
+ ],
+ "score": 1.0,
+ "content": "One possibility for learning a causal structure (through SCM modelling) is to perform many ex-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 198,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 198,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "periments in which one executes interventions. Thus, such interventions modify the effect of the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "intervened upon variable from its parents in the corresponding DAG, which the model has to quickly",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 233
+ ],
+ "score": 1.0,
+ "content": "adapt to. We can make parallel connections to meta-learning, where the inner loop can be consid-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "ered as fast adaptation to the distribution change, and outer loop can be considered as learning the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 242,
+ 505,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 505,
+ 255
+ ],
+ "score": 1.0,
+ "content": "stationary meta-parameters of the model. For causal induction, one can consider each distribution",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "which arises as a result of an intervention as a meta-example, and use a meta-learning objective for",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 277
+ ],
+ "score": 1.0,
+ "content": "fast adaptation in response to an intervention. One can think of model parameters as being composed",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 274,
+ 506,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 506,
+ 289
+ ],
+ "score": 1.0,
+ "content": "of slow- and fast-changing parameters. The slow parameters are analogous to the meta-parameters in",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 298
+ ],
+ "score": 1.0,
+ "content": "meta-learning and are used for (1) intervention prediction in order to handle the unknown intervention",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "and for (2) modeling the underlying causal structure. On the other hand, the fast parameters are used",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 308,
+ 505,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 505,
+ 321
+ ],
+ "score": 1.0,
+ "content": "to model the effect of interventions. An explicit search over the exponentially-growing space of all",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 505,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 505,
+ 331
+ ],
+ "score": 1.0,
+ "content": "possible DAGs is avoided by modeling the conditionals of the structural causal model using function",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 330,
+ 506,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 330,
+ 457,
+ 341
+ ],
+ "score": 1.0,
+ "content": "approximators, with one neural network per variable. The belief over whether one node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 458,
+ 330,
+ 462,
+ 339
+ ],
+ "score": 0.73,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 463,
+ 330,
+ 506,
+ 341
+ ],
+ "score": 1.0,
+ "content": "is a direct",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 341,
+ 504,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 341,
+ 227,
+ 353
+ ],
+ "score": 1.0,
+ "content": "causal parent of another node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 342,
+ 233,
+ 352
+ ],
+ "score": 0.81,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 234,
+ 341,
+ 411,
+ 353
+ ],
+ "score": 1.0,
+ "content": "corresponds to a dropout probability for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 342,
+ 416,
+ 351
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 416,
+ 341,
+ 497,
+ 353
+ ],
+ "score": 1.0,
+ "content": "-th input of network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 498,
+ 342,
+ 504,
+ 352
+ ],
+ "score": 0.77,
+ "content": "j",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 352,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 352,
+ 204,
+ 367
+ ],
+ "score": 1.0,
+ "content": "(which predicts variable",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 354,
+ 210,
+ 365
+ ],
+ "score": 0.81,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 352,
+ 392,
+ 367
+ ],
+ "score": 1.0,
+ "content": "given its parents). This cheaply represents all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 393,
+ 352,
+ 411,
+ 363
+ ],
+ "score": 0.89,
+ "content": "2 ^ { M ^ { 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 352,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "possible model graphs,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 506,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 506,
+ 378
+ ],
+ "score": 1.0,
+ "content": "with the graph search implicitly achieved by learning these dropout probabilities. We thus propose a",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 376,
+ 506,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 376,
+ 506,
+ 388
+ ],
+ "score": 1.0,
+ "content": "new method for fast adaptation and learning of neural causal models by framing the problem in a",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 387,
+ 413,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 387,
+ 413,
+ 399
+ ],
+ "score": 1.0,
+ "content": "meta-learning setting, similar to (Dasgupta et al., 2019; Bengio et al., 2019).",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 403,
+ 399,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 403,
+ 400,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 403,
+ 400,
+ 416
+ ],
+ "score": 1.0,
+ "content": "Our contributions Our key contributions can be summarized as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 132,
+ 424,
+ 505,
+ 521
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 424,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 424,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "• Handle causal induction to the case where the variable on which a soft intervention took",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 141,
+ 436,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 436,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "place is not known by the learner, and show that better results can be obtained when the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 141,
+ 447,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 447,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "learner attempts to infer that information and uses it to appropriately change parameters and",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 142,
+ 459,
+ 213,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 459,
+ 213,
+ 469
+ ],
+ "score": 1.0,
+ "content": "meta-parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 131,
+ 471,
+ 506,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 471,
+ 506,
+ 485
+ ],
+ "score": 1.0,
+ "content": "• We bypass the issue of having to optimize over and represent an exponentially large set of",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 142,
+ 484,
+ 470,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 484,
+ 470,
+ 495
+ ],
+ "score": 1.0,
+ "content": "discrete causal graphs by learning an efficiently parametrized ensemble of SCMs,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 131,
+ 498,
+ 505,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 498,
+ 505,
+ 510
+ ],
+ "score": 1.0,
+ "content": "• Show that our algorithm correctly identifies the causal graph and use the learned graph for",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 142,
+ 510,
+ 308,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 510,
+ 308,
+ 521
+ ],
+ "score": 1.0,
+ "content": "generalization to an unseen environment.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 536,
+ 208,
+ 549
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 535,
+ 209,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 535,
+ 209,
+ 551
+ ],
+ "score": 1.0,
+ "content": "2 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 560,
+ 504,
+ 584
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 410,
+ 573
+ ],
+ "score": 1.0,
+ "content": "A Structural Causal Model (SCM) (Peters et al., 2017) over a finite number",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 410,
+ 562,
+ 421,
+ 571
+ ],
+ "score": 0.85,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 560,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "of random variables",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 572,
+ 252,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 574,
+ 119,
+ 583
+ ],
+ "score": 0.87,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 572,
+ 252,
+ 585
+ ],
+ "score": 1.0,
+ "content": "is a set of structural assignments",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 205,
+ 589,
+ 405,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 205,
+ 589,
+ 405,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 205,
+ 589,
+ 405,
+ 604
+ ],
+ "score": 0.89,
+ "content": "X _ { i } : = f _ { i } ( X _ { p a ( i , C ) } , N _ { i } ) , \\quad \\forall i \\in \\{ 0 , \\dots , M - 1 \\}",
+ "type": "interline_equation",
+ "image_path": "216c3e36b771714960a490774f1f132f5450e31656442971f4f56bc6dd455527.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 40,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 205,
+ 589,
+ 405,
+ 604
+ ],
+ "spans": [],
+ "index": 40
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 608,
+ 505,
+ 686
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 506,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 133,
+ 622
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 610,
+ 146,
+ 620
+ ],
+ "score": 0.88,
+ "content": "N _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 609,
+ 280,
+ 622
+ ],
+ "score": 1.0,
+ "content": "is jointly-independent noise and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 609,
+ 315,
+ 621
+ ],
+ "score": 0.93,
+ "content": "p a ( i , C )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 609,
+ 506,
+ 622
+ ],
+ "score": 1.0,
+ "content": "is the set of parents (direct causes) of variable",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 107,
+ 617,
+ 507,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 622,
+ 111,
+ 630
+ ],
+ "score": 0.79,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 617,
+ 195,
+ 634
+ ],
+ "score": 1.0,
+ "content": "under configuration",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 621,
+ 204,
+ 630
+ ],
+ "score": 0.83,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 617,
+ 369,
+ 634
+ ],
+ "score": 1.0,
+ "content": "of the SCM directed acyclic graph, i.e.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 369,
+ 619,
+ 438,
+ 632
+ ],
+ "score": 0.94,
+ "content": "C \\in \\{ 0 , 1 \\} ^ { M \\times M }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 617,
+ 463,
+ 634
+ ],
+ "score": 1.0,
+ "content": ", with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 464,
+ 621,
+ 495,
+ 632
+ ],
+ "score": 0.91,
+ "content": "c _ { i j } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 617,
+ 507,
+ 634
+ ],
+ "score": 1.0,
+ "content": "if",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 104,
+ 630,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 630,
+ 128,
+ 645
+ ],
+ "score": 1.0,
+ "content": "node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 632,
+ 132,
+ 641
+ ],
+ "score": 0.78,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 630,
+ 171,
+ 645
+ ],
+ "score": 1.0,
+ "content": "has node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 632,
+ 177,
+ 643
+ ],
+ "score": 0.83,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 630,
+ 278,
+ 645
+ ],
+ "score": 1.0,
+ "content": "as a parent (equivalently,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 632,
+ 338,
+ 644
+ ],
+ "score": 0.92,
+ "content": "X _ { j } \\in X _ { p a ( i , C ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 630,
+ 357,
+ 645
+ ],
+ "score": 1.0,
+ "content": "; i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 632,
+ 371,
+ 643
+ ],
+ "score": 0.89,
+ "content": "X _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 630,
+ 448,
+ 645
+ ],
+ "score": 1.0,
+ "content": "is a direct cause of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 448,
+ 632,
+ 460,
+ 642
+ ],
+ "score": 0.87,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 460,
+ 630,
+ 486,
+ 645
+ ],
+ "score": 1.0,
+ "content": "). The",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 486,
+ 633,
+ 493,
+ 641
+ ],
+ "score": 0.8,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 630,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "-th",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 641,
+ 506,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 227,
+ 654
+ ],
+ "score": 1.0,
+ "content": "power of an adjacency matrix,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 642,
+ 241,
+ 652
+ ],
+ "score": 0.86,
+ "content": "C ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 242,
+ 641,
+ 359,
+ 654
+ ],
+ "score": 1.0,
+ "content": ", counts the number of length-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 645,
+ 366,
+ 652
+ ],
+ "score": 0.8,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 366,
+ 641,
+ 435,
+ 654
+ ],
+ "score": 1.0,
+ "content": "walks from node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 435,
+ 643,
+ 440,
+ 652
+ ],
+ "score": 0.79,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 440,
+ 641,
+ 473,
+ 654
+ ],
+ "score": 1.0,
+ "content": "to node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 643,
+ 479,
+ 654
+ ],
+ "score": 0.83,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 641,
+ 506,
+ 654
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 506,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 174,
+ 666
+ ],
+ "score": 1.0,
+ "content": "graph in element",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 655,
+ 187,
+ 666
+ ],
+ "score": 0.87,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 653,
+ 255,
+ 666
+ ],
+ "score": 1.0,
+ "content": ". The trace of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 255,
+ 655,
+ 262,
+ 663
+ ],
+ "score": 0.75,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 653,
+ 394,
+ 666
+ ],
+ "score": 1.0,
+ "content": "-th power of an adjacency matrix,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 395,
+ 653,
+ 426,
+ 665
+ ],
+ "score": 0.9,
+ "content": "\\mathrm { T r } ( C ^ { n } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 653,
+ 506,
+ 666
+ ],
+ "score": 1.0,
+ "content": ", counts the number",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 145,
+ 676
+ ],
+ "score": 1.0,
+ "content": "of length-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 666,
+ 153,
+ 674
+ ],
+ "score": 0.73,
+ "content": "\\mathbf { \\nabla } \\cdot n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 153,
+ 664,
+ 473,
+ 676
+ ],
+ "score": 1.0,
+ "content": "cycles in the graph. Causal structure learning is the recovery of the ground-truth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 665,
+ 482,
+ 674
+ ],
+ "score": 0.81,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 664,
+ 505,
+ 676
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 676,
+ 269,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 676,
+ 269,
+ 686
+ ],
+ "score": 1.0,
+ "content": "observational and interventional studies.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "Functional and structural meta-parameters Let us consider the simplest SCMs, those with",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 107,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 710,
+ 140,
+ 720
+ ],
+ "score": 0.89,
+ "content": "M = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 709,
+ 326,
+ 722
+ ],
+ "score": 1.0,
+ "content": "random variables (lets say random variables",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 711,
+ 335,
+ 720
+ ],
+ "score": 0.83,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 709,
+ 355,
+ 722
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 355,
+ 712,
+ 364,
+ 720
+ ],
+ "score": 0.83,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "). Only three DAGs exist relating",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 167,
+ 734
+ ],
+ "score": 1.0,
+ "content": "them; they are",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 721,
+ 200,
+ 731
+ ],
+ "score": 0.87,
+ "content": "A B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 719,
+ 204,
+ 734
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 721,
+ 237,
+ 730
+ ],
+ "score": 0.87,
+ "content": "B A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 719,
+ 250,
+ 734
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 721,
+ 274,
+ 730
+ ],
+ "score": 0.81,
+ "content": "A \\bot B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 719,
+ 466,
+ 734
+ ],
+ "score": 1.0,
+ "content": ". These can be represented as the following. If",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 721,
+ 475,
+ 730
+ ],
+ "score": 0.79,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 475,
+ 719,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "causes",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 181
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "However, in the real world, interventions are also not always performed by an experimenter. They",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "can be performed by other agents, or by environmental changes in ways that are unknown, or by",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "a naive learner (like a robot) which does not know precisely yet how its low-level actions change",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "high-level causal variables. In this paper, we look at the setting where interventions are unknown, and",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 126,
+ 507,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 126,
+ 507,
+ 140
+ ],
+ "score": 1.0,
+ "content": "our goal is to discover causal graphs given unknown-intervention samples. The challenging aspect of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 507,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 507,
+ 150
+ ],
+ "score": 1.0,
+ "content": "this setting is to not only learn the causal graph structure, but also predict the intervention accurately.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 147,
+ 506,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 147,
+ 506,
+ 161
+ ],
+ "score": 1.0,
+ "content": "In this setting, we need to make sure to: (1) avoid an exponential search over all possible DAGs, (2)",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 156,
+ 506,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 156,
+ 506,
+ 173
+ ],
+ "score": 1.0,
+ "content": "handle unknown interventions, (3) model the effect of interventions, and (4) model the underlying",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 170,
+ 173,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 170,
+ 173,
+ 182
+ ],
+ "score": 1.0,
+ "content": "causal structure.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 104,
+ 81,
+ 507,
+ 182
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 186,
+ 505,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 200
+ ],
+ "score": 1.0,
+ "content": "One possibility for learning a causal structure (through SCM modelling) is to perform many ex-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 198,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 198,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "periments in which one executes interventions. Thus, such interventions modify the effect of the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "intervened upon variable from its parents in the corresponding DAG, which the model has to quickly",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 219,
+ 506,
+ 233
+ ],
+ "score": 1.0,
+ "content": "adapt to. We can make parallel connections to meta-learning, where the inner loop can be consid-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "ered as fast adaptation to the distribution change, and outer loop can be considered as learning the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 242,
+ 505,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 505,
+ 255
+ ],
+ "score": 1.0,
+ "content": "stationary meta-parameters of the model. For causal induction, one can consider each distribution",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "which arises as a result of an intervention as a meta-example, and use a meta-learning objective for",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 277
+ ],
+ "score": 1.0,
+ "content": "fast adaptation in response to an intervention. One can think of model parameters as being composed",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 274,
+ 506,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 274,
+ 506,
+ 289
+ ],
+ "score": 1.0,
+ "content": "of slow- and fast-changing parameters. The slow parameters are analogous to the meta-parameters in",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 506,
+ 298
+ ],
+ "score": 1.0,
+ "content": "meta-learning and are used for (1) intervention prediction in order to handle the unknown intervention",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 297,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "and for (2) modeling the underlying causal structure. On the other hand, the fast parameters are used",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 308,
+ 505,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 505,
+ 321
+ ],
+ "score": 1.0,
+ "content": "to model the effect of interventions. An explicit search over the exponentially-growing space of all",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 505,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 505,
+ 331
+ ],
+ "score": 1.0,
+ "content": "possible DAGs is avoided by modeling the conditionals of the structural causal model using function",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 330,
+ 506,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 330,
+ 457,
+ 341
+ ],
+ "score": 1.0,
+ "content": "approximators, with one neural network per variable. The belief over whether one node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 458,
+ 330,
+ 462,
+ 339
+ ],
+ "score": 0.73,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 463,
+ 330,
+ 506,
+ 341
+ ],
+ "score": 1.0,
+ "content": "is a direct",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 341,
+ 504,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 341,
+ 227,
+ 353
+ ],
+ "score": 1.0,
+ "content": "causal parent of another node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 342,
+ 233,
+ 352
+ ],
+ "score": 0.81,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 234,
+ 341,
+ 411,
+ 353
+ ],
+ "score": 1.0,
+ "content": "corresponds to a dropout probability for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 342,
+ 416,
+ 351
+ ],
+ "score": 0.75,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 416,
+ 341,
+ 497,
+ 353
+ ],
+ "score": 1.0,
+ "content": "-th input of network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 498,
+ 342,
+ 504,
+ 352
+ ],
+ "score": 0.77,
+ "content": "j",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 352,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 352,
+ 204,
+ 367
+ ],
+ "score": 1.0,
+ "content": "(which predicts variable",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 354,
+ 210,
+ 365
+ ],
+ "score": 0.81,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 210,
+ 352,
+ 392,
+ 367
+ ],
+ "score": 1.0,
+ "content": "given its parents). This cheaply represents all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 393,
+ 352,
+ 411,
+ 363
+ ],
+ "score": 0.89,
+ "content": "2 ^ { M ^ { 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 352,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "possible model graphs,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 506,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 506,
+ 378
+ ],
+ "score": 1.0,
+ "content": "with the graph search implicitly achieved by learning these dropout probabilities. We thus propose a",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 376,
+ 506,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 376,
+ 506,
+ 388
+ ],
+ "score": 1.0,
+ "content": "new method for fast adaptation and learning of neural causal models by framing the problem in a",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 387,
+ 413,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 387,
+ 413,
+ 399
+ ],
+ "score": 1.0,
+ "content": "meta-learning setting, similar to (Dasgupta et al., 2019; Bengio et al., 2019).",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 104,
+ 187,
+ 506,
+ 399
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 403,
+ 399,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 403,
+ 400,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 403,
+ 400,
+ 416
+ ],
+ "score": 1.0,
+ "content": "Our contributions Our key contributions can be summarized as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 107,
+ 403,
+ 400,
+ 416
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 132,
+ 424,
+ 505,
+ 521
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 424,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 424,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "• Handle causal induction to the case where the variable on which a soft intervention took",
+ "type": "text"
+ }
+ ],
+ "index": 29,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 436,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 436,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "place is not known by the learner, and show that better results can be obtained when the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 141,
+ 447,
+ 506,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 447,
+ 506,
+ 460
+ ],
+ "score": 1.0,
+ "content": "learner attempts to infer that information and uses it to appropriately change parameters and",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 142,
+ 459,
+ 213,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 459,
+ 213,
+ 469
+ ],
+ "score": 1.0,
+ "content": "meta-parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 32,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 131,
+ 471,
+ 506,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 471,
+ 506,
+ 485
+ ],
+ "score": 1.0,
+ "content": "• We bypass the issue of having to optimize over and represent an exponentially large set of",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 142,
+ 484,
+ 470,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 484,
+ 470,
+ 495
+ ],
+ "score": 1.0,
+ "content": "discrete causal graphs by learning an efficiently parametrized ensemble of SCMs,",
+ "type": "text"
+ }
+ ],
+ "index": 34,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 131,
+ 498,
+ 505,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 498,
+ 505,
+ 510
+ ],
+ "score": 1.0,
+ "content": "• Show that our algorithm correctly identifies the causal graph and use the learned graph for",
+ "type": "text"
+ }
+ ],
+ "index": 35,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 142,
+ 510,
+ 308,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 510,
+ 308,
+ 521
+ ],
+ "score": 1.0,
+ "content": "generalization to an unseen environment.",
+ "type": "text"
+ }
+ ],
+ "index": 36,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 131,
+ 424,
+ 506,
+ 521
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 536,
+ 208,
+ 549
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 535,
+ 209,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 535,
+ 209,
+ 551
+ ],
+ "score": 1.0,
+ "content": "2 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 560,
+ 504,
+ 584
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 410,
+ 573
+ ],
+ "score": 1.0,
+ "content": "A Structural Causal Model (SCM) (Peters et al., 2017) over a finite number",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 410,
+ 562,
+ 421,
+ 571
+ ],
+ "score": 0.85,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 560,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "of random variables",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 572,
+ 252,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 574,
+ 119,
+ 583
+ ],
+ "score": 0.87,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 572,
+ 252,
+ 585
+ ],
+ "score": 1.0,
+ "content": "is a set of structural assignments",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 106,
+ 560,
+ 505,
+ 585
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 205,
+ 589,
+ 405,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 205,
+ 589,
+ 405,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 205,
+ 589,
+ 405,
+ 604
+ ],
+ "score": 0.89,
+ "content": "X _ { i } : = f _ { i } ( X _ { p a ( i , C ) } , N _ { i } ) , \\quad \\forall i \\in \\{ 0 , \\dots , M - 1 \\}",
+ "type": "interline_equation",
+ "image_path": "216c3e36b771714960a490774f1f132f5450e31656442971f4f56bc6dd455527.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 40,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 205,
+ 589,
+ 405,
+ 604
+ ],
+ "spans": [],
+ "index": 40
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 608,
+ 505,
+ 686
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 506,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 133,
+ 622
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 610,
+ 146,
+ 620
+ ],
+ "score": 0.88,
+ "content": "N _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 609,
+ 280,
+ 622
+ ],
+ "score": 1.0,
+ "content": "is jointly-independent noise and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 609,
+ 315,
+ 621
+ ],
+ "score": 0.93,
+ "content": "p a ( i , C )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 315,
+ 609,
+ 506,
+ 622
+ ],
+ "score": 1.0,
+ "content": "is the set of parents (direct causes) of variable",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 107,
+ 617,
+ 507,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 622,
+ 111,
+ 630
+ ],
+ "score": 0.79,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 111,
+ 617,
+ 195,
+ 634
+ ],
+ "score": 1.0,
+ "content": "under configuration",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 621,
+ 204,
+ 630
+ ],
+ "score": 0.83,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 617,
+ 369,
+ 634
+ ],
+ "score": 1.0,
+ "content": "of the SCM directed acyclic graph, i.e.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 369,
+ 619,
+ 438,
+ 632
+ ],
+ "score": 0.94,
+ "content": "C \\in \\{ 0 , 1 \\} ^ { M \\times M }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 617,
+ 463,
+ 634
+ ],
+ "score": 1.0,
+ "content": ", with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 464,
+ 621,
+ 495,
+ 632
+ ],
+ "score": 0.91,
+ "content": "c _ { i j } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 617,
+ 507,
+ 634
+ ],
+ "score": 1.0,
+ "content": "if",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 104,
+ 630,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 630,
+ 128,
+ 645
+ ],
+ "score": 1.0,
+ "content": "node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 632,
+ 132,
+ 641
+ ],
+ "score": 0.78,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 630,
+ 171,
+ 645
+ ],
+ "score": 1.0,
+ "content": "has node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 632,
+ 177,
+ 643
+ ],
+ "score": 0.83,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 630,
+ 278,
+ 645
+ ],
+ "score": 1.0,
+ "content": "as a parent (equivalently,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 632,
+ 338,
+ 644
+ ],
+ "score": 0.92,
+ "content": "X _ { j } \\in X _ { p a ( i , C ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 630,
+ 357,
+ 645
+ ],
+ "score": 1.0,
+ "content": "; i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 632,
+ 371,
+ 643
+ ],
+ "score": 0.89,
+ "content": "X _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 630,
+ 448,
+ 645
+ ],
+ "score": 1.0,
+ "content": "is a direct cause of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 448,
+ 632,
+ 460,
+ 642
+ ],
+ "score": 0.87,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 460,
+ 630,
+ 486,
+ 645
+ ],
+ "score": 1.0,
+ "content": "). The",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 486,
+ 633,
+ 493,
+ 641
+ ],
+ "score": 0.8,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 630,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "-th",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 641,
+ 506,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 227,
+ 654
+ ],
+ "score": 1.0,
+ "content": "power of an adjacency matrix,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 642,
+ 241,
+ 652
+ ],
+ "score": 0.86,
+ "content": "C ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 242,
+ 641,
+ 359,
+ 654
+ ],
+ "score": 1.0,
+ "content": ", counts the number of length-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 645,
+ 366,
+ 652
+ ],
+ "score": 0.8,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 366,
+ 641,
+ 435,
+ 654
+ ],
+ "score": 1.0,
+ "content": "walks from node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 435,
+ 643,
+ 440,
+ 652
+ ],
+ "score": 0.79,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 440,
+ 641,
+ 473,
+ 654
+ ],
+ "score": 1.0,
+ "content": "to node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 643,
+ 479,
+ 654
+ ],
+ "score": 0.83,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 641,
+ 506,
+ 654
+ ],
+ "score": 1.0,
+ "content": "of the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 506,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 174,
+ 666
+ ],
+ "score": 1.0,
+ "content": "graph in element",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 655,
+ 187,
+ 666
+ ],
+ "score": 0.87,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 653,
+ 255,
+ 666
+ ],
+ "score": 1.0,
+ "content": ". The trace of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 255,
+ 655,
+ 262,
+ 663
+ ],
+ "score": 0.75,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 653,
+ 394,
+ 666
+ ],
+ "score": 1.0,
+ "content": "-th power of an adjacency matrix,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 395,
+ 653,
+ 426,
+ 665
+ ],
+ "score": 0.9,
+ "content": "\\mathrm { T r } ( C ^ { n } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 653,
+ 506,
+ 666
+ ],
+ "score": 1.0,
+ "content": ", counts the number",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 664,
+ 505,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 145,
+ 676
+ ],
+ "score": 1.0,
+ "content": "of length-",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 666,
+ 153,
+ 674
+ ],
+ "score": 0.73,
+ "content": "\\mathbf { \\nabla } \\cdot n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 153,
+ 664,
+ 473,
+ 676
+ ],
+ "score": 1.0,
+ "content": "cycles in the graph. Causal structure learning is the recovery of the ground-truth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 665,
+ 482,
+ 674
+ ],
+ "score": 0.81,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 664,
+ 505,
+ 676
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 676,
+ 269,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 676,
+ 269,
+ 686
+ ],
+ "score": 1.0,
+ "content": "observational and interventional studies.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 44,
+ "bbox_fs": [
+ 104,
+ 609,
+ 507,
+ 686
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "Functional and structural meta-parameters Let us consider the simplest SCMs, those with",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 107,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 710,
+ 140,
+ 720
+ ],
+ "score": 0.89,
+ "content": "M = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 709,
+ 326,
+ 722
+ ],
+ "score": 1.0,
+ "content": "random variables (lets say random variables",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 711,
+ 335,
+ 720
+ ],
+ "score": 0.83,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 709,
+ 355,
+ 722
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 355,
+ 712,
+ 364,
+ 720
+ ],
+ "score": 0.83,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "). Only three DAGs exist relating",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 167,
+ 734
+ ],
+ "score": 1.0,
+ "content": "them; they are",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 168,
+ 721,
+ 200,
+ 731
+ ],
+ "score": 0.87,
+ "content": "A B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 719,
+ 204,
+ 734
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 721,
+ 237,
+ 730
+ ],
+ "score": 0.87,
+ "content": "B A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 719,
+ 250,
+ 734
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 721,
+ 274,
+ 730
+ ],
+ "score": 0.81,
+ "content": "A \\bot B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 719,
+ 466,
+ 734
+ ],
+ "score": 1.0,
+ "content": ". These can be represented as the following. If",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 721,
+ 475,
+ 730
+ ],
+ "score": 0.79,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 475,
+ 719,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "causes",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49,
+ "bbox_fs": [
+ 105,
+ 699,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 127
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 81,
+ 502,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 83,
+ 115,
+ 92
+ ],
+ "score": 0.79,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 81,
+ 215,
+ 96
+ ],
+ "score": 1.0,
+ "content": ", then the SCM between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 215,
+ 83,
+ 224,
+ 92
+ ],
+ "score": 0.73,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 224,
+ 81,
+ 242,
+ 96
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 83,
+ 252,
+ 93
+ ],
+ "score": 0.79,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 81,
+ 342,
+ 96
+ ],
+ "score": 1.0,
+ "content": "can be represented as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 342,
+ 82,
+ 434,
+ 95
+ ],
+ "score": 0.92,
+ "content": "B = f _ { \\theta _ { B } } ( c _ { B A } \\cdot A , \\epsilon _ { B } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 81,
+ 465,
+ 96
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 83,
+ 502,
+ 94
+ ],
+ "score": 0.9,
+ "content": "c _ { B A } = 1",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 94,
+ 505,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 94,
+ 146,
+ 105
+ ],
+ "score": 0.91,
+ "content": "\\epsilon _ { B } \\sim N _ { B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 147,
+ 94,
+ 328,
+ 105
+ ],
+ "score": 1.0,
+ "content": "is a sample of an independent noise factor. If",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 94,
+ 338,
+ 104
+ ],
+ "score": 0.8,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 94,
+ 367,
+ 105
+ ],
+ "score": 1.0,
+ "content": "causes",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 95,
+ 375,
+ 104
+ ],
+ "score": 0.78,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 94,
+ 505,
+ 105
+ ],
+ "score": 1.0,
+ "content": ", then the SCM can be presented",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 117,
+ 116
+ ],
+ "score": 1.0,
+ "content": "as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 118,
+ 105,
+ 209,
+ 116
+ ],
+ "score": 0.92,
+ "content": "A = f _ { \\theta _ { A } } ( c _ { A B } \\cdot B , \\epsilon _ { A } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 105,
+ 240,
+ 116
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 240,
+ 105,
+ 277,
+ 115
+ ],
+ "score": 0.88,
+ "content": "c _ { A B } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 105,
+ 280,
+ 116
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 105,
+ 320,
+ 116
+ ],
+ "score": 0.89,
+ "content": "\\epsilon _ { A } \\sim N _ { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 320,
+ 105,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": "is a sample of an independent noise factor. If",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 115,
+ 483,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 116,
+ 115,
+ 125
+ ],
+ "score": 0.79,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 115,
+ 133,
+ 128
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 117,
+ 143,
+ 125
+ ],
+ "score": 0.81,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 143,
+ 115,
+ 388,
+ 128
+ ],
+ "score": 1.0,
+ "content": "are independent, then we have the same equations but we set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 116,
+ 425,
+ 127
+ ],
+ "score": 0.91,
+ "content": "c _ { A B } = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 425,
+ 115,
+ 442,
+ 128
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 116,
+ 479,
+ 127
+ ],
+ "score": 0.91,
+ "content": "c _ { B A } = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 480,
+ 115,
+ 483,
+ 128
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 1.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 132,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "We may now think of learning the structural causal model as learning the two probabilities of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 107,
+ 142,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 146,
+ 124,
+ 154
+ ],
+ "score": 0.86,
+ "content": "c _ { A B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 125,
+ 142,
+ 138,
+ 156
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 138,
+ 145,
+ 156,
+ 155
+ ],
+ "score": 0.87,
+ "content": "c _ { B A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 157,
+ 142,
+ 477,
+ 156
+ ],
+ "score": 1.0,
+ "content": "being 1 (versus 0), representing our belief in the causal relationship between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 144,
+ 486,
+ 153
+ ],
+ "score": 0.76,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 142,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 153,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 155,
+ 116,
+ 164
+ ],
+ "score": 0.78,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 153,
+ 376,
+ 167
+ ],
+ "score": 1.0,
+ "content": ". We can parametrize these probabilities differentiably using",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 377,
+ 154,
+ 485,
+ 166
+ ],
+ "score": 0.9,
+ "content": "P ( c _ { A B } = 1 ) ^ { \\bar { } } = \\sigma ( \\gamma _ { A B } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 153,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 107,
+ 162,
+ 507,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 165,
+ 213,
+ 177
+ ],
+ "score": 0.92,
+ "content": "P ( c _ { B A } = 1 ) = \\sigma ( \\gamma _ { B A } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 162,
+ 240,
+ 183
+ ],
+ "score": 1.0,
+ "content": ", with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 240,
+ 167,
+ 248,
+ 177
+ ],
+ "score": 0.81,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 162,
+ 326,
+ 183
+ ],
+ "score": 1.0,
+ "content": "real numbers and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 165,
+ 390,
+ 179
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r } { \\sigma ( x ) \\ = \\ \\frac { \\bar { } 1 } { 1 + e ^ { - x } } } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 391,
+ 162,
+ 507,
+ 183
+ ],
+ "score": 1.0,
+ "content": ". We can also parametrize",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 202,
+ 191
+ ],
+ "score": 1.0,
+ "content": "the structural equations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 178,
+ 217,
+ 190
+ ],
+ "score": 0.91,
+ "content": "f _ { \\theta _ { A } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 177,
+ 236,
+ 191
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 178,
+ 252,
+ 190
+ ],
+ "score": 0.91,
+ "content": "f _ { \\theta _ { B } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 177,
+ 505,
+ 191
+ ],
+ "score": 1.0,
+ "content": "in a differentiable manner, using conditional probability tables",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 189,
+ 217,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 189,
+ 217,
+ 200
+ ],
+ "score": 1.0,
+ "content": "(CPTs) or neural networks.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 205,
+ 505,
+ 348
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 206,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 206,
+ 462,
+ 217
+ ],
+ "score": 1.0,
+ "content": "Hence, the problem becomes simultaneously learning the structural meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 462,
+ 207,
+ 470,
+ 217
+ ],
+ "score": 0.81,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 206,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "and the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 217,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 217,
+ 221,
+ 228
+ ],
+ "score": 1.0,
+ "content": "functional meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 221,
+ 217,
+ 227,
+ 226
+ ],
+ "score": 0.77,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 217,
+ 349,
+ 228
+ ],
+ "score": 1.0,
+ "content": ". Functional meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 349,
+ 217,
+ 355,
+ 226
+ ],
+ "score": 0.76,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 356,
+ 217,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "may be easily learned by maximum",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 227,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 360,
+ 240
+ ],
+ "score": 1.0,
+ "content": "likelihood or a proxy, and using backpropagation. However, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 229,
+ 368,
+ 239
+ ],
+ "score": 0.83,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 227,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "are more difficult to infer. Bengio",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 104,
+ 237,
+ 507,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 237,
+ 507,
+ 252
+ ],
+ "score": 1.0,
+ "content": "et al. (2019) propose to learn them by using observations of how the distribution changes sparsely,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 248,
+ 506,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 248,
+ 506,
+ 263
+ ],
+ "score": 1.0,
+ "content": "with the transfer generalization (the adaptation rate after the sparse change) being the training",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 160,
+ 272
+ ],
+ "score": 1.0,
+ "content": "objective for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 262,
+ 167,
+ 271
+ ],
+ "score": 0.8,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 260,
+ 296,
+ 272
+ ],
+ "score": 1.0,
+ "content": ". Simultaneously inferring both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 261,
+ 302,
+ 270
+ ],
+ "score": 0.8,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 260,
+ 320,
+ 272
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 262,
+ 328,
+ 272
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 328,
+ 260,
+ 432,
+ 272
+ ],
+ "score": 1.0,
+ "content": "is still more difficult. An",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 261,
+ 444,
+ 270
+ ],
+ "score": 0.76,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 444,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "-variable SCM",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 196,
+ 286
+ ],
+ "score": 1.0,
+ "content": "over random variables",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 271,
+ 209,
+ 282
+ ],
+ "score": 0.84,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 270,
+ 214,
+ 286
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 214,
+ 271,
+ 295,
+ 283
+ ],
+ "score": 0.91,
+ "content": "i \\in \\{ 0 , \\ldots , M - 1 \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 296,
+ 270,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "can induce a super-exponential number of adjacency",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 142,
+ 295
+ ],
+ "score": 1.0,
+ "content": "matrices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 282,
+ 151,
+ 292
+ ],
+ "score": 0.77,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 282,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": ". The super-exponential explosion in the number of potential graph connectivity patterns",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "and the super-exponentially growing storage requirements of their defining conditional probability",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 504,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 381,
+ 317
+ ],
+ "score": 1.0,
+ "content": "tables make CPT-based parametrizations of the structural assignments",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 382,
+ 304,
+ 391,
+ 316
+ ],
+ "score": 0.88,
+ "content": "f _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 303,
+ 492,
+ 317
+ ],
+ "score": 1.0,
+ "content": "increasingly unwieldy as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 492,
+ 305,
+ 504,
+ 314
+ ],
+ "score": 0.78,
+ "content": "M",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 314,
+ 506,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 314,
+ 297,
+ 329
+ ],
+ "score": 1.0,
+ "content": "scales. As shown below, neural networks with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 316,
+ 310,
+ 327
+ ],
+ "score": 0.87,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 314,
+ 506,
+ 329
+ ],
+ "score": 1.0,
+ "content": "-masked inputs can provide a more manageable",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "parametrization. For more background about different kinds of intervention we ask the reader to refer",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 337,
+ 168,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 337,
+ 168,
+ 349
+ ],
+ "score": 1.0,
+ "content": "Appendix A.3.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 361,
+ 481,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 360,
+ 482,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 360,
+ 482,
+ 376
+ ],
+ "score": 1.0,
+ "content": "3 PROPOSED FRAMEWORK: META LEARNING FOR CAUSAL INDUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 386,
+ 505,
+ 452
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 507,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 507,
+ 400
+ ],
+ "score": 1.0,
+ "content": "Our framework disentangles the slow-changing meta-parameters, which reflect the stationary prop-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 397,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 397,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "erties discovered by the learner, and the fast-changing parameters, which adapt in response to",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": "interventional changes in distribution. We consider two kinds of meta-parameters: the causal graph",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 419,
+ 505,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 419,
+ 145,
+ 432
+ ],
+ "score": 1.0,
+ "content": "structure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 421,
+ 153,
+ 431
+ ],
+ "score": 0.81,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 153,
+ 419,
+ 281,
+ 432
+ ],
+ "score": 1.0,
+ "content": "and the model’s slow weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 419,
+ 301,
+ 430
+ ],
+ "score": 0.91,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 301,
+ 419,
+ 505,
+ 432
+ ],
+ "score": 1.0,
+ "content": ", along with the meta-learning objective for both",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 428,
+ 443
+ ],
+ "score": 1.0,
+ "content": "of them. We also consider one kind of parameter: the model’s fast weights,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 428,
+ 430,
+ 445,
+ 442
+ ],
+ "score": 0.9,
+ "content": "\\theta _ { \\mathrm { f a s t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 429,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": ". We will call",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 107,
+ 440,
+ 506,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 441,
+ 171,
+ 452
+ ],
+ "score": 0.92,
+ "content": "\\theta = \\theta _ { \\mathrm { s l o w } } + \\theta _ { \\mathrm { f a s t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 440,
+ 506,
+ 454
+ ],
+ "score": 1.0,
+ "content": "the sum of the slow, stationary meta-parameters and the fast, adaptational parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 460,
+ 216,
+ 471
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 217,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 217,
+ 472
+ ],
+ "score": 1.0,
+ "content": "3.1 TASK DESCRIPTION",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 480,
+ 505,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 481,
+ 505,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 481,
+ 505,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Our task setup deviates from most common deep learning modeling setups. For the purposes of this",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": "work, we restrict ourselves to inference of randomly-generated or manually-provided ground-truth",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 146,
+ 515
+ ],
+ "score": 1.0,
+ "content": "SCMs of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 503,
+ 158,
+ 513
+ ],
+ "score": 0.73,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 159,
+ 502,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "categorical random variables causally related via a DAG. The model is permitted to",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "see (1) data from the original ground-truth model, and (2) data from a modified ground-truth model",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "with a random intervention applied. In our experiments, at most one intervention is concurrently",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 548
+ ],
+ "score": 1.0,
+ "content": "performed. When an intervention is performed, a single node is randomly and uniformly chosen",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 150,
+ 559
+ ],
+ "score": 1.0,
+ "content": "among all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 547,
+ 162,
+ 557
+ ],
+ "score": 0.67,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 547,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "nodes, and its ground-truth distribution soft-intervened upon. The learner model is",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "aware of the samples having come from an intervention distribution, but is not aware of the identity",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 581
+ ],
+ "score": 1.0,
+ "content": "of the intervention node, and so must predict it. Each run of sampling steps under a given intervention",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "is referred to as an episode. The learner, over a large number of episodes, will experience all nodes",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 448,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 448,
+ 603
+ ],
+ "score": 1.0,
+ "content": "being intervened upon, and should be able to infer the SCM from these interventions.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 615,
+ 358,
+ 626
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 360,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 360,
+ 628
+ ],
+ "score": 1.0,
+ "content": "3.2 CAUSAL INDUCTION AS AN OPTIMIZATION PROBLEM",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 635,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "We first explain how we mitigate the problem of searching in the super-exponential set of graph",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 505,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 204,
+ 660
+ ],
+ "score": 1.0,
+ "content": "structures. If there are",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 647,
+ 217,
+ 657
+ ],
+ "score": 0.69,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 217,
+ 646,
+ 505,
+ 660
+ ],
+ "score": 1.0,
+ "content": "such variables, the strategy of considering all the possible structural",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 104,
+ 657,
+ 507,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 657,
+ 438,
+ 673
+ ],
+ "score": 1.0,
+ "content": "graphs as separate hypotheses is not feasible because it would require maintaining",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 438,
+ 658,
+ 472,
+ 672
+ ],
+ "score": 0.93,
+ "content": "O ( 2 ^ { M ^ { 2 } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 472,
+ 657,
+ 507,
+ 673
+ ],
+ "score": 1.0,
+ "content": "models",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 104,
+ 671,
+ 505,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 671,
+ 352,
+ 686
+ ],
+ "score": 1.0,
+ "content": "of the data. We note that we can cheaply choose any of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 671,
+ 370,
+ 683
+ ],
+ "score": 0.89,
+ "content": "2 ^ { M ^ { 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 671,
+ 505,
+ 686
+ ],
+ "score": 1.0,
+ "content": "possible DAGs through suitable",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 683,
+ 506,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 683,
+ 227,
+ 698
+ ],
+ "score": 1.0,
+ "content": "independent Bernoulli choices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 686,
+ 240,
+ 696
+ ],
+ "score": 0.86,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 240,
+ 683,
+ 345,
+ 698
+ ],
+ "score": 1.0,
+ "content": "associated with each edge",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 685,
+ 370,
+ 695
+ ],
+ "score": 0.9,
+ "content": "i j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 370,
+ 683,
+ 506,
+ 698
+ ],
+ "score": 1.0,
+ "content": "of the causal graph, i.e., sampling",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 695,
+ 505,
+ 709
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 695,
+ 134,
+ 709
+ ],
+ "score": 1.0,
+ "content": "all the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 696,
+ 147,
+ 708
+ ],
+ "score": 0.86,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 147,
+ 695,
+ 347,
+ 709
+ ],
+ "score": 1.0,
+ "content": "’s independently. Then we only need to learn the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 696,
+ 363,
+ 706
+ ],
+ "score": 0.88,
+ "content": "M ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 695,
+ 414,
+ 709
+ ],
+ "score": 1.0,
+ "content": "coefficients",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 697,
+ 427,
+ 708
+ ],
+ "score": 0.88,
+ "content": "\\gamma _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 428,
+ 695,
+ 505,
+ 709
+ ],
+ "score": 1.0,
+ "content": ", and we implicitly",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 104,
+ 705,
+ 508,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 705,
+ 232,
+ 724
+ ],
+ "score": 1.0,
+ "content": "maintain a distribution over the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 708,
+ 250,
+ 720
+ ],
+ "score": 0.89,
+ "content": "2 ^ { M ^ { 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 705,
+ 449,
+ 724
+ ],
+ "score": 1.0,
+ "content": "models corresponding to all the possible draws of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 711,
+ 462,
+ 722
+ ],
+ "score": 0.87,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 705,
+ 508,
+ 724
+ ],
+ "score": 1.0,
+ "content": ". Note that",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 240,
+ 734
+ ],
+ "score": 1.0,
+ "content": "a slight dependency between the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 240,
+ 722,
+ 253,
+ 733
+ ],
+ "score": 0.88,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 721,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "is induced if we require the causal graph to be acyclic (which",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 46.5
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 127
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 81,
+ 502,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 83,
+ 115,
+ 92
+ ],
+ "score": 0.79,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 81,
+ 215,
+ 96
+ ],
+ "score": 1.0,
+ "content": ", then the SCM between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 215,
+ 83,
+ 224,
+ 92
+ ],
+ "score": 0.73,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 224,
+ 81,
+ 242,
+ 96
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 83,
+ 252,
+ 93
+ ],
+ "score": 0.79,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 81,
+ 342,
+ 96
+ ],
+ "score": 1.0,
+ "content": "can be represented as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 342,
+ 82,
+ 434,
+ 95
+ ],
+ "score": 0.92,
+ "content": "B = f _ { \\theta _ { B } } ( c _ { B A } \\cdot A , \\epsilon _ { B } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 81,
+ 465,
+ 96
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 466,
+ 83,
+ 502,
+ 94
+ ],
+ "score": 0.9,
+ "content": "c _ { B A } = 1",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 94,
+ 505,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 94,
+ 146,
+ 105
+ ],
+ "score": 0.91,
+ "content": "\\epsilon _ { B } \\sim N _ { B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 147,
+ 94,
+ 328,
+ 105
+ ],
+ "score": 1.0,
+ "content": "is a sample of an independent noise factor. If",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 94,
+ 338,
+ 104
+ ],
+ "score": 0.8,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 94,
+ 367,
+ 105
+ ],
+ "score": 1.0,
+ "content": "causes",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 95,
+ 375,
+ 104
+ ],
+ "score": 0.78,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 94,
+ 505,
+ 105
+ ],
+ "score": 1.0,
+ "content": ", then the SCM can be presented",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 117,
+ 116
+ ],
+ "score": 1.0,
+ "content": "as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 118,
+ 105,
+ 209,
+ 116
+ ],
+ "score": 0.92,
+ "content": "A = f _ { \\theta _ { A } } ( c _ { A B } \\cdot B , \\epsilon _ { A } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 105,
+ 240,
+ 116
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 240,
+ 105,
+ 277,
+ 115
+ ],
+ "score": 0.88,
+ "content": "c _ { A B } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 105,
+ 280,
+ 116
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 105,
+ 320,
+ 116
+ ],
+ "score": 0.89,
+ "content": "\\epsilon _ { A } \\sim N _ { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 320,
+ 105,
+ 506,
+ 116
+ ],
+ "score": 1.0,
+ "content": "is a sample of an independent noise factor. If",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 107,
+ 115,
+ 483,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 116,
+ 115,
+ 125
+ ],
+ "score": 0.79,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 115,
+ 133,
+ 128
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 117,
+ 143,
+ 125
+ ],
+ "score": 0.81,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 143,
+ 115,
+ 388,
+ 128
+ ],
+ "score": 1.0,
+ "content": "are independent, then we have the same equations but we set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 116,
+ 425,
+ 127
+ ],
+ "score": 0.91,
+ "content": "c _ { A B } = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 425,
+ 115,
+ 442,
+ 128
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 116,
+ 479,
+ 127
+ ],
+ "score": 0.91,
+ "content": "c _ { B A } = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 480,
+ 115,
+ 483,
+ 128
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 1.5,
+ "bbox_fs": [
+ 105,
+ 81,
+ 506,
+ 128
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 132,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "We may now think of learning the structural causal model as learning the two probabilities of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 107,
+ 142,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 146,
+ 124,
+ 154
+ ],
+ "score": 0.86,
+ "content": "c _ { A B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 125,
+ 142,
+ 138,
+ 156
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 138,
+ 145,
+ 156,
+ 155
+ ],
+ "score": 0.87,
+ "content": "c _ { B A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 157,
+ 142,
+ 477,
+ 156
+ ],
+ "score": 1.0,
+ "content": "being 1 (versus 0), representing our belief in the causal relationship between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 144,
+ 486,
+ 153
+ ],
+ "score": 0.76,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 142,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 153,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 155,
+ 116,
+ 164
+ ],
+ "score": 0.78,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 153,
+ 376,
+ 167
+ ],
+ "score": 1.0,
+ "content": ". We can parametrize these probabilities differentiably using",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 377,
+ 154,
+ 485,
+ 166
+ ],
+ "score": 0.9,
+ "content": "P ( c _ { A B } = 1 ) ^ { \\bar { } } = \\sigma ( \\gamma _ { A B } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 153,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 107,
+ 162,
+ 507,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 165,
+ 213,
+ 177
+ ],
+ "score": 0.92,
+ "content": "P ( c _ { B A } = 1 ) = \\sigma ( \\gamma _ { B A } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 162,
+ 240,
+ 183
+ ],
+ "score": 1.0,
+ "content": ", with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 240,
+ 167,
+ 248,
+ 177
+ ],
+ "score": 0.81,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 162,
+ 326,
+ 183
+ ],
+ "score": 1.0,
+ "content": "real numbers and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 165,
+ 390,
+ 179
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r } { \\sigma ( x ) \\ = \\ \\frac { \\bar { } 1 } { 1 + e ^ { - x } } } \\end{array}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 391,
+ 162,
+ 507,
+ 183
+ ],
+ "score": 1.0,
+ "content": ". We can also parametrize",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 202,
+ 191
+ ],
+ "score": 1.0,
+ "content": "the structural equations",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 202,
+ 178,
+ 217,
+ 190
+ ],
+ "score": 0.91,
+ "content": "f _ { \\theta _ { A } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 177,
+ 236,
+ 191
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 178,
+ 252,
+ 190
+ ],
+ "score": 0.91,
+ "content": "f _ { \\theta _ { B } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 177,
+ 505,
+ 191
+ ],
+ "score": 1.0,
+ "content": "in a differentiable manner, using conditional probability tables",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 189,
+ 217,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 189,
+ 217,
+ 200
+ ],
+ "score": 1.0,
+ "content": "(CPTs) or neural networks.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 132,
+ 507,
+ 200
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 205,
+ 505,
+ 348
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 206,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 206,
+ 462,
+ 217
+ ],
+ "score": 1.0,
+ "content": "Hence, the problem becomes simultaneously learning the structural meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 462,
+ 207,
+ 470,
+ 217
+ ],
+ "score": 0.81,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 206,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "and the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 217,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 217,
+ 221,
+ 228
+ ],
+ "score": 1.0,
+ "content": "functional meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 221,
+ 217,
+ 227,
+ 226
+ ],
+ "score": 0.77,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 217,
+ 349,
+ 228
+ ],
+ "score": 1.0,
+ "content": ". Functional meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 349,
+ 217,
+ 355,
+ 226
+ ],
+ "score": 0.76,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 356,
+ 217,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "may be easily learned by maximum",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 227,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 227,
+ 360,
+ 240
+ ],
+ "score": 1.0,
+ "content": "likelihood or a proxy, and using backpropagation. However, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 229,
+ 368,
+ 239
+ ],
+ "score": 0.83,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 227,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "are more difficult to infer. Bengio",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 104,
+ 237,
+ 507,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 237,
+ 507,
+ 252
+ ],
+ "score": 1.0,
+ "content": "et al. (2019) propose to learn them by using observations of how the distribution changes sparsely,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 248,
+ 506,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 248,
+ 506,
+ 263
+ ],
+ "score": 1.0,
+ "content": "with the transfer generalization (the adaptation rate after the sparse change) being the training",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 160,
+ 272
+ ],
+ "score": 1.0,
+ "content": "objective for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 262,
+ 167,
+ 271
+ ],
+ "score": 0.8,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 260,
+ 296,
+ 272
+ ],
+ "score": 1.0,
+ "content": ". Simultaneously inferring both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 261,
+ 302,
+ 270
+ ],
+ "score": 0.8,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 260,
+ 320,
+ 272
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 262,
+ 328,
+ 272
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 328,
+ 260,
+ 432,
+ 272
+ ],
+ "score": 1.0,
+ "content": "is still more difficult. An",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 261,
+ 444,
+ 270
+ ],
+ "score": 0.76,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 444,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "-variable SCM",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 196,
+ 286
+ ],
+ "score": 1.0,
+ "content": "over random variables",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 271,
+ 209,
+ 282
+ ],
+ "score": 0.84,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 270,
+ 214,
+ 286
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 214,
+ 271,
+ 295,
+ 283
+ ],
+ "score": 0.91,
+ "content": "i \\in \\{ 0 , \\ldots , M - 1 \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 296,
+ 270,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "can induce a super-exponential number of adjacency",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 142,
+ 295
+ ],
+ "score": 1.0,
+ "content": "matrices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 282,
+ 151,
+ 292
+ ],
+ "score": 0.77,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 282,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": ". The super-exponential explosion in the number of potential graph connectivity patterns",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "and the super-exponentially growing storage requirements of their defining conditional probability",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 504,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 381,
+ 317
+ ],
+ "score": 1.0,
+ "content": "tables make CPT-based parametrizations of the structural assignments",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 382,
+ 304,
+ 391,
+ 316
+ ],
+ "score": 0.88,
+ "content": "f _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 303,
+ 492,
+ 317
+ ],
+ "score": 1.0,
+ "content": "increasingly unwieldy as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 492,
+ 305,
+ 504,
+ 314
+ ],
+ "score": 0.78,
+ "content": "M",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 314,
+ 506,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 314,
+ 297,
+ 329
+ ],
+ "score": 1.0,
+ "content": "scales. As shown below, neural networks with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 316,
+ 310,
+ 327
+ ],
+ "score": 0.87,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 314,
+ 506,
+ 329
+ ],
+ "score": 1.0,
+ "content": "-masked inputs can provide a more manageable",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "parametrization. For more background about different kinds of intervention we ask the reader to refer",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 337,
+ 168,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 337,
+ 168,
+ 349
+ ],
+ "score": 1.0,
+ "content": "Appendix A.3.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 104,
+ 206,
+ 507,
+ 349
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 361,
+ 481,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 360,
+ 482,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 360,
+ 482,
+ 376
+ ],
+ "score": 1.0,
+ "content": "3 PROPOSED FRAMEWORK: META LEARNING FOR CAUSAL INDUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 386,
+ 505,
+ 452
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 507,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 507,
+ 400
+ ],
+ "score": 1.0,
+ "content": "Our framework disentangles the slow-changing meta-parameters, which reflect the stationary prop-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 397,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 397,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "erties discovered by the learner, and the fast-changing parameters, which adapt in response to",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": "interventional changes in distribution. We consider two kinds of meta-parameters: the causal graph",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 419,
+ 505,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 419,
+ 145,
+ 432
+ ],
+ "score": 1.0,
+ "content": "structure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 421,
+ 153,
+ 431
+ ],
+ "score": 0.81,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 153,
+ 419,
+ 281,
+ 432
+ ],
+ "score": 1.0,
+ "content": "and the model’s slow weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 419,
+ 301,
+ 430
+ ],
+ "score": 0.91,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 301,
+ 419,
+ 505,
+ 432
+ ],
+ "score": 1.0,
+ "content": ", along with the meta-learning objective for both",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 428,
+ 443
+ ],
+ "score": 1.0,
+ "content": "of them. We also consider one kind of parameter: the model’s fast weights,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 428,
+ 430,
+ 445,
+ 442
+ ],
+ "score": 0.9,
+ "content": "\\theta _ { \\mathrm { f a s t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 429,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": ". We will call",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 107,
+ 440,
+ 506,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 441,
+ 171,
+ 452
+ ],
+ "score": 0.92,
+ "content": "\\theta = \\theta _ { \\mathrm { s l o w } } + \\theta _ { \\mathrm { f a s t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 440,
+ 506,
+ 454
+ ],
+ "score": 1.0,
+ "content": "the sum of the slow, stationary meta-parameters and the fast, adaptational parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 385,
+ 507,
+ 454
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 460,
+ 216,
+ 471
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 217,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 217,
+ 472
+ ],
+ "score": 1.0,
+ "content": "3.1 TASK DESCRIPTION",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 480,
+ 505,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 481,
+ 505,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 481,
+ 505,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Our task setup deviates from most common deep learning modeling setups. For the purposes of this",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": "work, we restrict ourselves to inference of randomly-generated or manually-provided ground-truth",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 146,
+ 515
+ ],
+ "score": 1.0,
+ "content": "SCMs of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 503,
+ 158,
+ 513
+ ],
+ "score": 0.73,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 159,
+ 502,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "categorical random variables causally related via a DAG. The model is permitted to",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "see (1) data from the original ground-truth model, and (2) data from a modified ground-truth model",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "with a random intervention applied. In our experiments, at most one intervention is concurrently",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 548
+ ],
+ "score": 1.0,
+ "content": "performed. When an intervention is performed, a single node is randomly and uniformly chosen",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 150,
+ 559
+ ],
+ "score": 1.0,
+ "content": "among all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 547,
+ 162,
+ 557
+ ],
+ "score": 0.67,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 547,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "nodes, and its ground-truth distribution soft-intervened upon. The learner model is",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "aware of the samples having come from an intervention distribution, but is not aware of the identity",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 581
+ ],
+ "score": 1.0,
+ "content": "of the intervention node, and so must predict it. Each run of sampling steps under a given intervention",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "is referred to as an episode. The learner, over a large number of episodes, will experience all nodes",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 448,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 448,
+ 603
+ ],
+ "score": 1.0,
+ "content": "being intervened upon, and should be able to infer the SCM from these interventions.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 481,
+ 506,
+ 603
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 615,
+ 358,
+ 626
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 360,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 360,
+ 628
+ ],
+ "score": 1.0,
+ "content": "3.2 CAUSAL INDUCTION AS AN OPTIMIZATION PROBLEM",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 635,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "We first explain how we mitigate the problem of searching in the super-exponential set of graph",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 505,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 204,
+ 660
+ ],
+ "score": 1.0,
+ "content": "structures. If there are",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 647,
+ 217,
+ 657
+ ],
+ "score": 0.69,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 217,
+ 646,
+ 505,
+ 660
+ ],
+ "score": 1.0,
+ "content": "such variables, the strategy of considering all the possible structural",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 104,
+ 657,
+ 507,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 657,
+ 438,
+ 673
+ ],
+ "score": 1.0,
+ "content": "graphs as separate hypotheses is not feasible because it would require maintaining",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 438,
+ 658,
+ 472,
+ 672
+ ],
+ "score": 0.93,
+ "content": "O ( 2 ^ { M ^ { 2 } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 472,
+ 657,
+ 507,
+ 673
+ ],
+ "score": 1.0,
+ "content": "models",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 104,
+ 671,
+ 505,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 671,
+ 352,
+ 686
+ ],
+ "score": 1.0,
+ "content": "of the data. We note that we can cheaply choose any of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 671,
+ 370,
+ 683
+ ],
+ "score": 0.89,
+ "content": "2 ^ { M ^ { 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 671,
+ 505,
+ 686
+ ],
+ "score": 1.0,
+ "content": "possible DAGs through suitable",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 683,
+ 506,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 683,
+ 227,
+ 698
+ ],
+ "score": 1.0,
+ "content": "independent Bernoulli choices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 686,
+ 240,
+ 696
+ ],
+ "score": 0.86,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 240,
+ 683,
+ 345,
+ 698
+ ],
+ "score": 1.0,
+ "content": "associated with each edge",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 685,
+ 370,
+ 695
+ ],
+ "score": 0.9,
+ "content": "i j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 370,
+ 683,
+ 506,
+ 698
+ ],
+ "score": 1.0,
+ "content": "of the causal graph, i.e., sampling",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 695,
+ 505,
+ 709
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 695,
+ 134,
+ 709
+ ],
+ "score": 1.0,
+ "content": "all the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 696,
+ 147,
+ 708
+ ],
+ "score": 0.86,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 147,
+ 695,
+ 347,
+ 709
+ ],
+ "score": 1.0,
+ "content": "’s independently. Then we only need to learn the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 696,
+ 363,
+ 706
+ ],
+ "score": 0.88,
+ "content": "M ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 695,
+ 414,
+ 709
+ ],
+ "score": 1.0,
+ "content": "coefficients",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 697,
+ 427,
+ 708
+ ],
+ "score": 0.88,
+ "content": "\\gamma _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 428,
+ 695,
+ 505,
+ 709
+ ],
+ "score": 1.0,
+ "content": ", and we implicitly",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 104,
+ 705,
+ 508,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 705,
+ 232,
+ 724
+ ],
+ "score": 1.0,
+ "content": "maintain a distribution over the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 708,
+ 250,
+ 720
+ ],
+ "score": 0.89,
+ "content": "2 ^ { M ^ { 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 705,
+ 449,
+ 724
+ ],
+ "score": 1.0,
+ "content": "models corresponding to all the possible draws of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 711,
+ 462,
+ 722
+ ],
+ "score": 0.87,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 705,
+ 508,
+ 724
+ ],
+ "score": 1.0,
+ "content": ". Note that",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 240,
+ 734
+ ],
+ "score": 1.0,
+ "content": "a slight dependency between the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 240,
+ 722,
+ 253,
+ 733
+ ],
+ "score": 0.88,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 721,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "is induced if we require the causal graph to be acyclic (which",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 205,
+ 95
+ ],
+ "score": 1.0,
+ "content": "allows one to sample the",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 205,
+ 83,
+ 216,
+ 93
+ ],
+ "score": 0.78,
+ "content": "X",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 216,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "using ancestral sampling). To enforce that constraint it is not sufficient to",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 137,
+ 106
+ ],
+ "score": 1.0,
+ "content": "require",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 137,
+ 94,
+ 179,
+ 106
+ ],
+ "score": 0.91,
+ "content": "c _ { i j } c _ { j i } = 0",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 180,
+ 93,
+ 489,
+ 106
+ ],
+ "score": 1.0,
+ "content": "(both cannot be 1). We deal with this problem with a regularizer acting on the",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 490,
+ 95,
+ 497,
+ 105
+ ],
+ "score": 0.78,
+ "content": "\\gamma",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 498,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "’s",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 332,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 332,
+ 117
+ ],
+ "score": 1.0,
+ "content": "in order to favour acyclic solutions (Zheng et al., 2018).",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 104,
+ 635,
+ 508,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 205,
+ 95
+ ],
+ "score": 1.0,
+ "content": "allows one to sample the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 83,
+ 216,
+ 93
+ ],
+ "score": 0.78,
+ "content": "X",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "using ancestral sampling). To enforce that constraint it is not sufficient to",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 137,
+ 106
+ ],
+ "score": 1.0,
+ "content": "require",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 137,
+ 94,
+ 179,
+ 106
+ ],
+ "score": 0.91,
+ "content": "c _ { i j } c _ { j i } = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 180,
+ 93,
+ 489,
+ 106
+ ],
+ "score": 1.0,
+ "content": "(both cannot be 1). We deal with this problem with a regularizer acting on the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 95,
+ 497,
+ 105
+ ],
+ "score": 0.78,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 498,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "’s",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 332,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 332,
+ 117
+ ],
+ "score": 1.0,
+ "content": "in order to favour acyclic solutions (Zheng et al., 2018).",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 119,
+ 258,
+ 329
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 259,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 259,
+ 130
+ ],
+ "score": 1.0,
+ "content": "In our approach, each random vari-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 129,
+ 259,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 259,
+ 141
+ ],
+ "score": 1.0,
+ "content": "able’s structural assignment is mod-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 139,
+ 257,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 139,
+ 142,
+ 154
+ ],
+ "score": 1.0,
+ "content": "eled via",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 140,
+ 257,
+ 153
+ ],
+ "score": 0.88,
+ "content": "X _ { i } : = f _ { \\theta _ { i } } ( \\bar { c } _ { i 0 } \\times X _ { 0 } , c _ { i 1 } \\times",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 153,
+ 258,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 154,
+ 196,
+ 165
+ ],
+ "score": 0.9,
+ "content": "X _ { 1 } , . . . , c _ { i m } \\times X _ { m } , \\epsilon _ { i } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 153,
+ 226,
+ 166
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 226,
+ 153,
+ 247,
+ 166
+ ],
+ "score": 0.91,
+ "content": "f _ { \\theta _ { i } } ( \\dot { ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 153,
+ 258,
+ 166
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 258,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 258,
+ 177
+ ],
+ "score": 1.0,
+ "content": "a neural network (MLP) with parame-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 174,
+ 259,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 174,
+ 123,
+ 189
+ ],
+ "score": 1.0,
+ "content": "ters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 123,
+ 176,
+ 132,
+ 187
+ ],
+ "score": 0.86,
+ "content": "\\theta _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 174,
+ 153,
+ 189
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 153,
+ 176,
+ 254,
+ 188
+ ],
+ "score": 0.89,
+ "content": "c _ { i j } \\sim \\mathrm { B i n } ( \\mathrm { s i g m o i d } ( \\gamma _ { i j } ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 174,
+ 259,
+ 189
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 185,
+ 258,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 185,
+ 258,
+ 199
+ ],
+ "score": 1.0,
+ "content": "Through this construction we can",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 198,
+ 257,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 257,
+ 209
+ ],
+ "score": 1.0,
+ "content": "frame the causal induction problem",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 256,
+ 220
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 250,
+ 220
+ ],
+ "score": 1.0,
+ "content": "as an optimization problem, with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 209,
+ 256,
+ 218
+ ],
+ "score": 0.71,
+ "content": "\\theta",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 220,
+ 258,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 220,
+ 258,
+ 230
+ ],
+ "score": 1.0,
+ "content": "optimized to maximize the likelihood",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 258,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 227,
+ 242
+ ],
+ "score": 1.0,
+ "content": "of data under the model but",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 232,
+ 236,
+ 242
+ ],
+ "score": 0.81,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 230,
+ 258,
+ 242
+ ],
+ "score": 1.0,
+ "content": "opti-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 241,
+ 258,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 258,
+ 254
+ ],
+ "score": 1.0,
+ "content": "mized with respect to a meta-learning",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 253,
+ 257,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 257,
+ 264
+ ],
+ "score": 1.0,
+ "content": "objective arising from changes in",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 263,
+ 259,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 263,
+ 259,
+ 275
+ ],
+ "score": 1.0,
+ "content": "distribution because of interventions.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 273,
+ 258,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 258,
+ 286
+ ],
+ "score": 1.0,
+ "content": "There are a few benefits for learning",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 259,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 259,
+ 296
+ ],
+ "score": 1.0,
+ "content": "a parametrized ensemble of SCMs.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 296,
+ 258,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 296,
+ 258,
+ 308
+ ],
+ "score": 1.0,
+ "content": "Such an ensemble is analogous to",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 258,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 258,
+ 320
+ ],
+ "score": 1.0,
+ "content": "an ensemble of neural nets differing",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 317,
+ 259,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 317,
+ 259,
+ 331
+ ],
+ "score": 1.0,
+ "content": "by their binary input dropout masks,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 262
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 262
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 262
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "d47a155d4b265e10c72c13a432ae0b74490eabb21b453e8a46d108a4b05b1c95.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 26,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 147.33333333333334
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 270,
+ 147.33333333333334,
+ 498,
+ 161.66666666666669
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 270,
+ 161.66666666666669,
+ 498,
+ 176.00000000000003
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 270,
+ 176.00000000000003,
+ 498,
+ 190.33333333333337
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 270,
+ 190.33333333333337,
+ 498,
+ 204.6666666666667
+ ],
+ "spans": [],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 270,
+ 204.6666666666667,
+ 498,
+ 219.00000000000006
+ ],
+ "spans": [],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 270,
+ 219.00000000000006,
+ 498,
+ 233.3333333333334
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 270,
+ 233.3333333333334,
+ 498,
+ 247.66666666666674
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 270,
+ 247.66666666666674,
+ 498,
+ 262.00000000000006
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 264,
+ 274,
+ 504,
+ 316
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 265,
+ 274,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 265,
+ 274,
+ 376,
+ 283
+ ],
+ "score": 1.0,
+ "content": "Figure 1: MLP Model Architecture for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 275,
+ 400,
+ 282
+ ],
+ "score": 0.87,
+ "content": "M = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 400,
+ 274,
+ 403,
+ 283
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 274,
+ 426,
+ 282
+ ],
+ "score": 0.86,
+ "content": "N = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 274,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "(fork3) SCM. The model",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 264,
+ 282,
+ 505,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 264,
+ 282,
+ 383,
+ 292
+ ],
+ "score": 1.0,
+ "content": "computes the conditional probabilities of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 282,
+ 410,
+ 291
+ ],
+ "score": 0.9,
+ "content": "{ \\hat { A } } , { \\hat { B } } , { \\hat { C } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 282,
+ 505,
+ 292
+ ],
+ "score": 1.0,
+ "content": "given their parents using a stack",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 264,
+ 291,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 264,
+ 291,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "of three independent MLPs. The MLP input layer uses an adjacency matrix sampled",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 265,
+ 299,
+ 505,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 265,
+ 300,
+ 281,
+ 308
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 299,
+ 316,
+ 308
+ ],
+ "score": 0.89,
+ "content": "\\mathrm { B e r } ( \\sigma ( \\gamma ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 300,
+ 505,
+ 308
+ ],
+ "score": 1.0,
+ "content": "as an input mask to force the model to make use only of parent",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 264,
+ 308,
+ 358,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 264,
+ 308,
+ 358,
+ 316
+ ],
+ "score": 1.0,
+ "content": "nodes to predict their child node.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 330,
+ 385,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 387,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 387,
+ 341
+ ],
+ "score": 1.0,
+ "content": "which select what variables are used as predictors of another variable.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 356,
+ 304,
+ 367
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 305,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 305,
+ 369
+ ],
+ "score": 1.0,
+ "content": "3.3 FAST ADAPTATION BY META-LEARNING",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 506,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 506,
+ 390
+ ],
+ "score": 1.0,
+ "content": "Fast and slow weights To disentangle an environment’s stable, unchanging properties (the causal",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 388,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 482,
+ 402
+ ],
+ "score": 1.0,
+ "content": "structure) from unstable, changing properties (the effects of an intervention), we proposed in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 389,
+ 493,
+ 400
+ ],
+ "score": 0.78,
+ "content": "\\ S 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 494,
+ 388,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 351,
+ 412
+ ],
+ "score": 1.0,
+ "content": "distinguish between the model’s functional meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 400,
+ 371,
+ 411
+ ],
+ "score": 0.91,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 399,
+ 435,
+ 412
+ ],
+ "score": 1.0,
+ "content": "and parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 400,
+ 452,
+ 411
+ ],
+ "score": 0.9,
+ "content": "\\theta _ { \\mathrm { f a s t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 399,
+ 506,
+ 412
+ ],
+ "score": 1.0,
+ "content": ". The sum of",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 409,
+ 503,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 164,
+ 424
+ ],
+ "score": 1.0,
+ "content": "these weights,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 411,
+ 227,
+ 422
+ ],
+ "score": 0.93,
+ "content": "\\theta = \\theta _ { \\mathrm { s l o w } } + \\theta _ { \\mathrm { f a s t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 228,
+ 409,
+ 432,
+ 424
+ ],
+ "score": 1.0,
+ "content": ", parametrizes the MLPs computing the conditionals",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 410,
+ 503,
+ 423
+ ],
+ "score": 0.94,
+ "content": "P _ { i } ( X _ { i } | X _ { p a ( i ) } ; \\theta _ { i } )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 506,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 506,
+ 434
+ ],
+ "score": 1.0,
+ "content": "The fast weights and the slow weights terminology is drawn from Hinton & Plaut (1987). The",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 505,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 170,
+ 445
+ ],
+ "score": 1.0,
+ "content": "construction of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 170,
+ 433,
+ 176,
+ 442
+ ],
+ "score": 0.8,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 432,
+ 505,
+ 445
+ ],
+ "score": 1.0,
+ "content": "as a sum of initial, slow weights plus zeroed, fast weight that are then allowed to",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "quickly adapt during a transfer episode is due to MAML (Finn et al., 2017). The ability to generalize",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 455,
+ 506,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 506,
+ 466
+ ],
+ "score": 1.0,
+ "content": "out-of-distribution by adapting to a transfer distribution can then be measured by the likelihood after",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 466,
+ 275,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 466,
+ 275,
+ 478
+ ],
+ "score": 1.0,
+ "content": "adapting the fast weights on transfer data.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 482,
+ 505,
+ 537
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "Since an intervention is generally not persistent from one transfer distribution to another, the model’s",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 201,
+ 506
+ ],
+ "score": 1.0,
+ "content": "functional parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 201,
+ 494,
+ 222,
+ 505
+ ],
+ "score": 0.84,
+ "content": "( \\theta _ { \\mathrm { f a s t } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 222,
+ 493,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "are reset after each episode of transfer distribution adaptation. The",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 177,
+ 517
+ ],
+ "score": 1.0,
+ "content": "meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 505,
+ 212,
+ 516
+ ],
+ "score": 0.86,
+ "content": "( \\theta _ { \\mathrm { { s l o w } } } , \\gamma )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 504,
+ 506,
+ 517
+ ],
+ "score": 1.0,
+ "content": "are preserved, then updated after each episode. Inspired by Bengio et al.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 514,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 459,
+ 528
+ ],
+ "score": 1.0,
+ "content": "(2019), the meta-objective for each meta-example over some intervention distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 460,
+ 515,
+ 479,
+ 527
+ ],
+ "score": 0.9,
+ "content": "D _ { \\mathrm { i n t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 514,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 236,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 236,
+ 539
+ ],
+ "score": 1.0,
+ "content": "following1 \"meta-transfer\" loss:",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 202,
+ 545,
+ 409,
+ 571
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 202,
+ 545,
+ 409,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 202,
+ 545,
+ 409,
+ 571
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { R } = - \\mathbb { E } _ { X \\sim D _ { \\mathrm { i n t } } } [ \\log \\mathbb { E } _ { C \\sim \\mathrm { B e r } ( \\gamma ) } [ \\prod _ { i } \\mathcal { L } _ { C , i } ( X ; \\theta _ { \\mathrm { s l o w } } ) ] ]",
+ "type": "interline_equation",
+ "image_path": "068e6a8e648f3a826e74fc8ffaa9b259768beb6cc68fd5adcb940044beeb4554.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 52,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 202,
+ 545,
+ 409,
+ 571
+ ],
+ "spans": [],
+ "index": 52
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 578,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 578,
+ 133,
+ 591
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 579,
+ 144,
+ 588
+ ],
+ "score": 0.84,
+ "content": "X",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 578,
+ 378,
+ 591
+ ],
+ "score": 1.0,
+ "content": "is an example sampled from the intervention distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 378,
+ 579,
+ 397,
+ 589
+ ],
+ "score": 0.86,
+ "content": "D _ { \\mathrm { i n t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 397,
+ 578,
+ 401,
+ 591
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 579,
+ 410,
+ 588
+ ],
+ "score": 0.75,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 578,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "is an adjacency matrix",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 106,
+ 590,
+ 490,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 590,
+ 315,
+ 601
+ ],
+ "score": 1.0,
+ "content": "drawn from our belief distribution (parametrized by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 591,
+ 322,
+ 601
+ ],
+ "score": 0.75,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 590,
+ 490,
+ 601
+ ],
+ "score": 1.0,
+ "content": ") about graph structure configurations and",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ }
+ ],
+ "index": 53.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 236,
+ 608,
+ 375,
+ 623
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 236,
+ 608,
+ 375,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 236,
+ 608,
+ 375,
+ 623
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { L } _ { C , i } ( X ) = P ( X _ { i } | X _ { p a ( i , C ) } ; \\theta _ { \\mathrm { s l o w } } )",
+ "type": "interline_equation",
+ "image_path": "546c50a94ce4f3eb42178e4df94a5efda19506263a40169ffcd706ef0871bded.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 55,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 236,
+ 608,
+ 375,
+ 623
+ ],
+ "spans": [],
+ "index": 55
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 628,
+ 504,
+ 653
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 198,
+ 642
+ ],
+ "score": 1.0,
+ "content": "is the likelihood of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 630,
+ 203,
+ 639
+ ],
+ "score": 0.76,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 628,
+ 250,
+ 642
+ ],
+ "score": 1.0,
+ "content": "-th variable",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 630,
+ 262,
+ 640
+ ],
+ "score": 0.89,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 628,
+ 320,
+ 642
+ ],
+ "score": 1.0,
+ "content": "of the sample",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 630,
+ 330,
+ 639
+ ],
+ "score": 0.85,
+ "content": "X",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 628,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": ", when predicting it under the configuration",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 107,
+ 636,
+ 298,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 641,
+ 115,
+ 650
+ ],
+ "score": 0.79,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 636,
+ 258,
+ 656
+ ],
+ "score": 1.0,
+ "content": "from the set of its putative parents,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 641,
+ 293,
+ 653
+ ],
+ "score": 0.92,
+ "content": "X _ { p a ( i , C ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 293,
+ 636,
+ 298,
+ 656
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ }
+ ],
+ "index": 56.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 665,
+ 505,
+ 699
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "Structural Parameter Gradient Estimator Because a discrete Bernoulli random sampling pro-",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "cess is used to produce the configurations under which the log-likelihood of data samples is obtained,",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 507,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 384,
+ 701
+ ],
+ "score": 1.0,
+ "content": "we require a gradient estimator to propagate gradient through to the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 385,
+ 691,
+ 392,
+ 699
+ ],
+ "score": 0.72,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 687,
+ 507,
+ 701
+ ],
+ "score": 1.0,
+ "content": "structural meta-parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ }
+ ],
+ "index": 59
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 105,
+ 711,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 709,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 709,
+ 339,
+ 723
+ ],
+ "score": 1.0,
+ "content": "1 These equations differ from Bengio et al. (2019) in that the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 339,
+ 713,
+ 343,
+ 720
+ ],
+ "score": 0.8,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 709,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "indices and products were dropped, because",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 215,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 215,
+ 732
+ ],
+ "score": 1.0,
+ "content": "the computation is not online.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": []
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [],
+ "index": 1,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 117
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 119,
+ 258,
+ 329
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 259,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 259,
+ 130
+ ],
+ "score": 1.0,
+ "content": "In our approach, each random vari-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 129,
+ 259,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 129,
+ 259,
+ 141
+ ],
+ "score": 1.0,
+ "content": "able’s structural assignment is mod-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 139,
+ 257,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 139,
+ 142,
+ 154
+ ],
+ "score": 1.0,
+ "content": "eled via",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 140,
+ 257,
+ 153
+ ],
+ "score": 0.88,
+ "content": "X _ { i } : = f _ { \\theta _ { i } } ( \\bar { c } _ { i 0 } \\times X _ { 0 } , c _ { i 1 } \\times",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 153,
+ 258,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 154,
+ 196,
+ 165
+ ],
+ "score": 0.9,
+ "content": "X _ { 1 } , . . . , c _ { i m } \\times X _ { m } , \\epsilon _ { i } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 196,
+ 153,
+ 226,
+ 166
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 226,
+ 153,
+ 247,
+ 166
+ ],
+ "score": 0.91,
+ "content": "f _ { \\theta _ { i } } ( \\dot { ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 153,
+ 258,
+ 166
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 258,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 258,
+ 177
+ ],
+ "score": 1.0,
+ "content": "a neural network (MLP) with parame-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 174,
+ 259,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 174,
+ 123,
+ 189
+ ],
+ "score": 1.0,
+ "content": "ters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 123,
+ 176,
+ 132,
+ 187
+ ],
+ "score": 0.86,
+ "content": "\\theta _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 174,
+ 153,
+ 189
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 153,
+ 176,
+ 254,
+ 188
+ ],
+ "score": 0.89,
+ "content": "c _ { i j } \\sim \\mathrm { B i n } ( \\mathrm { s i g m o i d } ( \\gamma _ { i j } ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 174,
+ 259,
+ 189
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 185,
+ 258,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 185,
+ 258,
+ 199
+ ],
+ "score": 1.0,
+ "content": "Through this construction we can",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 198,
+ 257,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 257,
+ 209
+ ],
+ "score": 1.0,
+ "content": "frame the causal induction problem",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 256,
+ 220
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 250,
+ 220
+ ],
+ "score": 1.0,
+ "content": "as an optimization problem, with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 209,
+ 256,
+ 218
+ ],
+ "score": 0.71,
+ "content": "\\theta",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 220,
+ 258,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 220,
+ 258,
+ 230
+ ],
+ "score": 1.0,
+ "content": "optimized to maximize the likelihood",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 258,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 227,
+ 242
+ ],
+ "score": 1.0,
+ "content": "of data under the model but",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 232,
+ 236,
+ 242
+ ],
+ "score": 0.81,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 230,
+ 258,
+ 242
+ ],
+ "score": 1.0,
+ "content": "opti-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 241,
+ 258,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 258,
+ 254
+ ],
+ "score": 1.0,
+ "content": "mized with respect to a meta-learning",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 253,
+ 257,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 257,
+ 264
+ ],
+ "score": 1.0,
+ "content": "objective arising from changes in",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 263,
+ 259,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 263,
+ 259,
+ 275
+ ],
+ "score": 1.0,
+ "content": "distribution because of interventions.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 273,
+ 258,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 273,
+ 258,
+ 286
+ ],
+ "score": 1.0,
+ "content": "There are a few benefits for learning",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 259,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 259,
+ 296
+ ],
+ "score": 1.0,
+ "content": "a parametrized ensemble of SCMs.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 296,
+ 258,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 296,
+ 258,
+ 308
+ ],
+ "score": 1.0,
+ "content": "Such an ensemble is analogous to",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 258,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 258,
+ 320
+ ],
+ "score": 1.0,
+ "content": "an ensemble of neural nets differing",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 317,
+ 259,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 317,
+ 259,
+ 331
+ ],
+ "score": 1.0,
+ "content": "by their binary input dropout masks,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 329,
+ 387,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 387,
+ 341
+ ],
+ "score": 1.0,
+ "content": "which select what variables are used as predictors of another variable.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 118,
+ 259,
+ 331
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 262
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 262
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 262
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "d47a155d4b265e10c72c13a432ae0b74490eabb21b453e8a46d108a4b05b1c95.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 26,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 270,
+ 133,
+ 498,
+ 147.33333333333334
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 270,
+ 147.33333333333334,
+ 498,
+ 161.66666666666669
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 270,
+ 161.66666666666669,
+ 498,
+ 176.00000000000003
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 270,
+ 176.00000000000003,
+ 498,
+ 190.33333333333337
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 270,
+ 190.33333333333337,
+ 498,
+ 204.6666666666667
+ ],
+ "spans": [],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 270,
+ 204.6666666666667,
+ 498,
+ 219.00000000000006
+ ],
+ "spans": [],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 270,
+ 219.00000000000006,
+ 498,
+ 233.3333333333334
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 270,
+ 233.3333333333334,
+ 498,
+ 247.66666666666674
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 270,
+ 247.66666666666674,
+ 498,
+ 262.00000000000006
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 264,
+ 274,
+ 504,
+ 316
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 265,
+ 274,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 265,
+ 274,
+ 376,
+ 283
+ ],
+ "score": 1.0,
+ "content": "Figure 1: MLP Model Architecture for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 275,
+ 400,
+ 282
+ ],
+ "score": 0.87,
+ "content": "M = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 400,
+ 274,
+ 403,
+ 283
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 403,
+ 274,
+ 426,
+ 282
+ ],
+ "score": 0.86,
+ "content": "N = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 274,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "(fork3) SCM. The model",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 264,
+ 282,
+ 505,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 264,
+ 282,
+ 383,
+ 292
+ ],
+ "score": 1.0,
+ "content": "computes the conditional probabilities of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 282,
+ 410,
+ 291
+ ],
+ "score": 0.9,
+ "content": "{ \\hat { A } } , { \\hat { B } } , { \\hat { C } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 282,
+ 505,
+ 292
+ ],
+ "score": 1.0,
+ "content": "given their parents using a stack",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 264,
+ 291,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 264,
+ 291,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "of three independent MLPs. The MLP input layer uses an adjacency matrix sampled",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 265,
+ 299,
+ 505,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 265,
+ 300,
+ 281,
+ 308
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 299,
+ 316,
+ 308
+ ],
+ "score": 0.89,
+ "content": "\\mathrm { B e r } ( \\sigma ( \\gamma ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 300,
+ 505,
+ 308
+ ],
+ "score": 1.0,
+ "content": "as an input mask to force the model to make use only of parent",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 264,
+ 308,
+ 358,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 264,
+ 308,
+ 358,
+ 316
+ ],
+ "score": 1.0,
+ "content": "nodes to predict their child node.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 330,
+ 385,
+ 340
+ ],
+ "lines": [],
+ "index": 36,
+ "bbox_fs": [
+ 106,
+ 329,
+ 387,
+ 341
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 356,
+ 304,
+ 367
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 305,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 356,
+ 305,
+ 369
+ ],
+ "score": 1.0,
+ "content": "3.3 FAST ADAPTATION BY META-LEARNING",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 506,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 506,
+ 390
+ ],
+ "score": 1.0,
+ "content": "Fast and slow weights To disentangle an environment’s stable, unchanging properties (the causal",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 388,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 482,
+ 402
+ ],
+ "score": 1.0,
+ "content": "structure) from unstable, changing properties (the effects of an intervention), we proposed in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 389,
+ 493,
+ 400
+ ],
+ "score": 0.78,
+ "content": "\\ S 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 494,
+ 388,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 351,
+ 412
+ ],
+ "score": 1.0,
+ "content": "distinguish between the model’s functional meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 400,
+ 371,
+ 411
+ ],
+ "score": 0.91,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 399,
+ 435,
+ 412
+ ],
+ "score": 1.0,
+ "content": "and parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 400,
+ 452,
+ 411
+ ],
+ "score": 0.9,
+ "content": "\\theta _ { \\mathrm { f a s t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 399,
+ 506,
+ 412
+ ],
+ "score": 1.0,
+ "content": ". The sum of",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 409,
+ 503,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 164,
+ 424
+ ],
+ "score": 1.0,
+ "content": "these weights,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 411,
+ 227,
+ 422
+ ],
+ "score": 0.93,
+ "content": "\\theta = \\theta _ { \\mathrm { s l o w } } + \\theta _ { \\mathrm { f a s t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 228,
+ 409,
+ 432,
+ 424
+ ],
+ "score": 1.0,
+ "content": ", parametrizes the MLPs computing the conditionals",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 410,
+ 503,
+ 423
+ ],
+ "score": 0.94,
+ "content": "P _ { i } ( X _ { i } | X _ { p a ( i ) } ; \\theta _ { i } )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 506,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 506,
+ 434
+ ],
+ "score": 1.0,
+ "content": "The fast weights and the slow weights terminology is drawn from Hinton & Plaut (1987). The",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 505,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 170,
+ 445
+ ],
+ "score": 1.0,
+ "content": "construction of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 170,
+ 433,
+ 176,
+ 442
+ ],
+ "score": 0.8,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 432,
+ 505,
+ 445
+ ],
+ "score": 1.0,
+ "content": "as a sum of initial, slow weights plus zeroed, fast weight that are then allowed to",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "quickly adapt during a transfer episode is due to MAML (Finn et al., 2017). The ability to generalize",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 455,
+ 506,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 506,
+ 466
+ ],
+ "score": 1.0,
+ "content": "out-of-distribution by adapting to a transfer distribution can then be measured by the likelihood after",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 466,
+ 275,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 466,
+ 275,
+ 478
+ ],
+ "score": 1.0,
+ "content": "adapting the fast weights on transfer data.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 105,
+ 377,
+ 506,
+ 478
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 482,
+ 505,
+ 537
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "Since an intervention is generally not persistent from one transfer distribution to another, the model’s",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 201,
+ 506
+ ],
+ "score": 1.0,
+ "content": "functional parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 201,
+ 494,
+ 222,
+ 505
+ ],
+ "score": 0.84,
+ "content": "( \\theta _ { \\mathrm { f a s t } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 222,
+ 493,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "are reset after each episode of transfer distribution adaptation. The",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 177,
+ 517
+ ],
+ "score": 1.0,
+ "content": "meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 505,
+ 212,
+ 516
+ ],
+ "score": 0.86,
+ "content": "( \\theta _ { \\mathrm { { s l o w } } } , \\gamma )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 504,
+ 506,
+ 517
+ ],
+ "score": 1.0,
+ "content": "are preserved, then updated after each episode. Inspired by Bengio et al.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 514,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 514,
+ 459,
+ 528
+ ],
+ "score": 1.0,
+ "content": "(2019), the meta-objective for each meta-example over some intervention distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 460,
+ 515,
+ 479,
+ 527
+ ],
+ "score": 0.9,
+ "content": "D _ { \\mathrm { i n t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 514,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "is the",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 236,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 236,
+ 539
+ ],
+ "score": 1.0,
+ "content": "following1 \"meta-transfer\" loss:",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49,
+ "bbox_fs": [
+ 105,
+ 483,
+ 506,
+ 539
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 202,
+ 545,
+ 409,
+ 571
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 202,
+ 545,
+ 409,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 202,
+ 545,
+ 409,
+ 571
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { R } = - \\mathbb { E } _ { X \\sim D _ { \\mathrm { i n t } } } [ \\log \\mathbb { E } _ { C \\sim \\mathrm { B e r } ( \\gamma ) } [ \\prod _ { i } \\mathcal { L } _ { C , i } ( X ; \\theta _ { \\mathrm { s l o w } } ) ] ]",
+ "type": "interline_equation",
+ "image_path": "068e6a8e648f3a826e74fc8ffaa9b259768beb6cc68fd5adcb940044beeb4554.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 52,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 202,
+ 545,
+ 409,
+ 571
+ ],
+ "spans": [],
+ "index": 52
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 578,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 578,
+ 133,
+ 591
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 579,
+ 144,
+ 588
+ ],
+ "score": 0.84,
+ "content": "X",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 578,
+ 378,
+ 591
+ ],
+ "score": 1.0,
+ "content": "is an example sampled from the intervention distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 378,
+ 579,
+ 397,
+ 589
+ ],
+ "score": 0.86,
+ "content": "D _ { \\mathrm { i n t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 397,
+ 578,
+ 401,
+ 591
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 579,
+ 410,
+ 588
+ ],
+ "score": 0.75,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 411,
+ 578,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "is an adjacency matrix",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 106,
+ 590,
+ 490,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 590,
+ 315,
+ 601
+ ],
+ "score": 1.0,
+ "content": "drawn from our belief distribution (parametrized by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 591,
+ 322,
+ 601
+ ],
+ "score": 0.75,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 590,
+ 490,
+ 601
+ ],
+ "score": 1.0,
+ "content": ") about graph structure configurations and",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ }
+ ],
+ "index": 53.5,
+ "bbox_fs": [
+ 106,
+ 578,
+ 505,
+ 601
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 236,
+ 608,
+ 375,
+ 623
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 236,
+ 608,
+ 375,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 236,
+ 608,
+ 375,
+ 623
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { L } _ { C , i } ( X ) = P ( X _ { i } | X _ { p a ( i , C ) } ; \\theta _ { \\mathrm { s l o w } } )",
+ "type": "interline_equation",
+ "image_path": "546c50a94ce4f3eb42178e4df94a5efda19506263a40169ffcd706ef0871bded.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 55,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 236,
+ 608,
+ 375,
+ 623
+ ],
+ "spans": [],
+ "index": 55
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 628,
+ 504,
+ 653
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 198,
+ 642
+ ],
+ "score": 1.0,
+ "content": "is the likelihood of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 630,
+ 203,
+ 639
+ ],
+ "score": 0.76,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 628,
+ 250,
+ 642
+ ],
+ "score": 1.0,
+ "content": "-th variable",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 630,
+ 262,
+ 640
+ ],
+ "score": 0.89,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 628,
+ 320,
+ 642
+ ],
+ "score": 1.0,
+ "content": "of the sample",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 630,
+ 330,
+ 639
+ ],
+ "score": 0.85,
+ "content": "X",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 628,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": ", when predicting it under the configuration",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 107,
+ 636,
+ 298,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 641,
+ 115,
+ 650
+ ],
+ "score": 0.79,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 116,
+ 636,
+ 258,
+ 656
+ ],
+ "score": 1.0,
+ "content": "from the set of its putative parents,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 641,
+ 293,
+ 653
+ ],
+ "score": 0.92,
+ "content": "X _ { p a ( i , C ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 293,
+ 636,
+ 298,
+ 656
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ }
+ ],
+ "index": 56.5,
+ "bbox_fs": [
+ 105,
+ 628,
+ 505,
+ 656
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 665,
+ 505,
+ 699
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "Structural Parameter Gradient Estimator Because a discrete Bernoulli random sampling pro-",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "cess is used to produce the configurations under which the log-likelihood of data samples is obtained,",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 507,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 384,
+ 701
+ ],
+ "score": 1.0,
+ "content": "we require a gradient estimator to propagate gradient through to the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 385,
+ 691,
+ 392,
+ 699
+ ],
+ "score": 0.72,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 687,
+ 507,
+ 701
+ ],
+ "score": 1.0,
+ "content": "structural meta-parameters.",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ }
+ ],
+ "index": 59,
+ "bbox_fs": [
+ 105,
+ 665,
+ 507,
+ 701
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 463,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 465,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 465,
+ 97
+ ],
+ "score": 1.0,
+ "content": "We adopt for this purpose the gradient estimate of Bengio et al. (2019) (but see footnote 1):",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 233,
+ 97,
+ 377,
+ 130
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 233,
+ 97,
+ 377,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 233,
+ 97,
+ 377,
+ 130
+ ],
+ "score": 0.95,
+ "content": "g _ { i j } = \\frac { \\sum _ { k } ( \\sigma ( \\gamma _ { i j } ) - c _ { i j } ^ { ( k ) } ) \\mathcal { L } _ { C , i } ^ { ( k ) } ( X ) } { \\sum _ { k } \\mathcal { L } _ { C , i } ^ { ( k ) } ( X ) }",
+ "type": "interline_equation",
+ "image_path": "05663c44b32bacf4e94ac7b3c8468e7e7e7174553d7cf2a1c4855974bd17f26a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 233,
+ 97,
+ 377,
+ 113.5
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 233,
+ 113.5,
+ 377,
+ 130.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 135,
+ 503,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 150,
+ 149
+ ],
+ "score": 1.0,
+ "content": "where the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 134,
+ 162,
+ 145
+ ],
+ "score": 0.53,
+ "content": "( k )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 132,
+ 366,
+ 149
+ ],
+ "score": 1.0,
+ "content": "superscript indicates the values obtained for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 136,
+ 373,
+ 145
+ ],
+ "score": 0.83,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 132,
+ 421,
+ 149
+ ],
+ "score": 1.0,
+ "content": "-th draw of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 422,
+ 136,
+ 430,
+ 145
+ ],
+ "score": 0.82,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 431,
+ 132,
+ 506,
+ 149
+ ],
+ "score": 1.0,
+ "content": ". This gradient is",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 145,
+ 455,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 194,
+ 159
+ ],
+ "score": 1.0,
+ "content": "estimated solely with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 147,
+ 213,
+ 157
+ ],
+ "score": 0.91,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 145,
+ 333,
+ 159
+ ],
+ "score": 1.0,
+ "content": "because estimates employing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 333,
+ 147,
+ 339,
+ 156
+ ],
+ "score": 0.79,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 145,
+ 455,
+ 159
+ ],
+ "score": 1.0,
+ "content": "have much greater variance.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 189
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "Acyclic Constraint We add a regularization term to the loss term that discourages the model from",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 423,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 423,
+ 191
+ ],
+ "score": 1.0,
+ "content": "having length-2 cycles in the learned adjacency matrix. The regularizer term is",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 186,
+ 193,
+ 425,
+ 221
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 186,
+ 193,
+ 425,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 186,
+ 193,
+ 425,
+ 221
+ ],
+ "score": 0.94,
+ "content": "J _ { \\mathrm { D A G } } = \\sum _ { i \\neq j } \\cosh ( \\sigma ( \\gamma _ { i j } ) \\sigma ( \\gamma _ { j i } ) ) , \\quad \\forall i , j \\in \\{ 0 , \\ldots , M - 1 \\}",
+ "type": "interline_equation",
+ "image_path": "c9e6c645bf54bf607c6162b9ac69233dcfc2d763372ff5553590f369bc6a63c0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 186,
+ 193,
+ 425,
+ 207.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 186,
+ 207.0,
+ 425,
+ 221.0
+ ],
+ "spans": [],
+ "index": 8
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 225,
+ 465,
+ 237
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 466,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 466,
+ 238
+ ],
+ "score": 1.0,
+ "content": "and is derived from Zheng et al. (2018). The details of the derivation are in the Appendix.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 250,
+ 226,
+ 261
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 249,
+ 227,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 249,
+ 227,
+ 262
+ ],
+ "score": 1.0,
+ "content": "3.4 MODEL DESCRIPTION",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 270,
+ 506,
+ 370
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 241,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Learner Model We model the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 271,
+ 253,
+ 280
+ ],
+ "score": 0.78,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 270,
+ 348,
+ 284
+ ],
+ "score": 1.0,
+ "content": "structural assignments",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 349,
+ 270,
+ 446,
+ 283
+ ],
+ "score": 0.92,
+ "content": "X _ { i } : = f _ { i } ( X _ { p a ( i , C ) } , N _ { i } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 270,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "(Eq. 1) of the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 281,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 281,
+ 169,
+ 293
+ ],
+ "score": 1.0,
+ "content": "SCM as a set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 282,
+ 181,
+ 291
+ ],
+ "score": 0.79,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 281,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "multi-layer perceptrons (MLPs), as in Bengio et al. (2019). The MLPs are identical",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 292,
+ 507,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 292,
+ 507,
+ 305
+ ],
+ "score": 1.0,
+ "content": "in shape but do not share any parameters, since they are modeling independent causal mechanisms.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 236,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Each possesses an input layer of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 303,
+ 269,
+ 313
+ ],
+ "score": 0.91,
+ "content": "M \\times N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 269,
+ 303,
+ 321,
+ 316
+ ],
+ "score": 1.0,
+ "content": "neurons (for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 304,
+ 333,
+ 313
+ ],
+ "score": 0.82,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 303,
+ 434,
+ 316
+ ],
+ "score": 1.0,
+ "content": "one-hot vectors of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 435,
+ 304,
+ 445,
+ 313
+ ],
+ "score": 0.79,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 303,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "each), a single",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 506,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 270,
+ 327
+ ],
+ "score": 1.0,
+ "content": "hidden layer chosen arbitrarily to have",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 270,
+ 314,
+ 331,
+ 326
+ ],
+ "score": 0.9,
+ "content": "\\operatorname* { m a x } ( 4 M , 4 N )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 313,
+ 506,
+ 327
+ ],
+ "score": 1.0,
+ "content": "neurons with a LeakyReLU activation of",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 237,
+ 338
+ ],
+ "score": 1.0,
+ "content": "slope 0.1, and an output layer of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 326,
+ 248,
+ 335
+ ],
+ "score": 0.84,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 325,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "neurons representing the unnormalized log-probabilities of each",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 504,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 180,
+ 349
+ ],
+ "score": 1.0,
+ "content": "category. To force",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 336,
+ 189,
+ 348
+ ],
+ "score": 0.87,
+ "content": "f _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 335,
+ 364,
+ 349
+ ],
+ "score": 1.0,
+ "content": "to rely exclusively on the direct ancestor set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 336,
+ 399,
+ 348
+ ],
+ "score": 0.92,
+ "content": "p a ( i , C )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 335,
+ 495,
+ 349
+ ],
+ "score": 1.0,
+ "content": "under adjacency matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 336,
+ 504,
+ 346
+ ],
+ "score": 0.79,
+ "content": "C",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 347,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 347,
+ 258,
+ 359
+ ],
+ "score": 1.0,
+ "content": "(See Eqn. 2), the one-hot input vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 348,
+ 272,
+ 359
+ ],
+ "score": 0.89,
+ "content": "X _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 347,
+ 322,
+ 359
+ ],
+ "score": 1.0,
+ "content": "for variable",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 347,
+ 335,
+ 358
+ ],
+ "score": 0.87,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 347,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": "’s MLP is masked by the Boolean element",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 357,
+ 462,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 119,
+ 370
+ ],
+ "score": 0.86,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 357,
+ 322,
+ 372
+ ],
+ "score": 1.0,
+ "content": ". The functional parameters of the MLP are the set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 358,
+ 459,
+ 370
+ ],
+ "score": 0.89,
+ "content": "\\theta = \\left\\{ { \\cal W } 0 _ { i h j n } , { \\tt B } 0 _ { i h } , { \\tt W } 1 _ { i n h } , { \\tt B } 1 _ { i n } \\right\\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 459,
+ 357,
+ 462,
+ 372
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 374,
+ 504,
+ 397
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 298,
+ 387
+ ],
+ "score": 1.0,
+ "content": "An example of the multi-MLP architecture with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 298,
+ 375,
+ 328,
+ 385
+ ],
+ "score": 0.9,
+ "content": "M = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 374,
+ 424,
+ 387
+ ],
+ "score": 1.0,
+ "content": "categorical variables of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 375,
+ 452,
+ 385
+ ],
+ "score": 0.9,
+ "content": "N = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 374,
+ 506,
+ 387
+ ],
+ "score": 1.0,
+ "content": "categories is",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 386,
+ 184,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 386,
+ 184,
+ 398
+ ],
+ "score": 1.0,
+ "content": "shown in Figure 1.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 409,
+ 505,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 409,
+ 506,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 409,
+ 506,
+ 422
+ ],
+ "score": 1.0,
+ "content": "Ground-Truth Model In our experiments, ground-truth SCM models are parametrized as a set of",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 420,
+ 505,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 420,
+ 505,
+ 432
+ ],
+ "score": 1.0,
+ "content": "MLPs of the same size as the learner models, thus avoid having to manually define the Conditional",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 444
+ ],
+ "score": 1.0,
+ "content": "Probability Tables (CPTs). Ground-truth models exist in two variants that differ mainly in initial-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 222,
+ 455
+ ],
+ "score": 1.0,
+ "content": "ization: synthetic, where the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 223,
+ 443,
+ 229,
+ 452
+ ],
+ "score": 0.77,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 442,
+ 360,
+ 455
+ ],
+ "score": 1.0,
+ "content": "are randomly initialized and the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 444,
+ 368,
+ 453
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 442,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "are either randomly-initialized or",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 452,
+ 506,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 273,
+ 466
+ ],
+ "score": 1.0,
+ "content": "pre-specified; and real-world, where the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 273,
+ 453,
+ 280,
+ 463
+ ],
+ "score": 0.78,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 280,
+ 452,
+ 298,
+ 466
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 455,
+ 306,
+ 465
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 452,
+ 506,
+ 466
+ ],
+ "score": 1.0,
+ "content": "are both initialized so as to closely replicate the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 465,
+ 246,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 246,
+ 476
+ ],
+ "score": 1.0,
+ "content": "CPTs of given Bayesian networks.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 488,
+ 202,
+ 500
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 203,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 203,
+ 501
+ ],
+ "score": 1.0,
+ "content": "3.5 INTERVENTIONS",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 509,
+ 505,
+ 554
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 504,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 352,
+ 522
+ ],
+ "score": 1.0,
+ "content": "Soft interventions To execute an intervention on variable",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 510,
+ 364,
+ 520
+ ],
+ "score": 0.88,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 509,
+ 430,
+ 522
+ ],
+ "score": 1.0,
+ "content": ", we reinitialize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 510,
+ 442,
+ 520
+ ],
+ "score": 0.87,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 443,
+ 509,
+ 504,
+ 522
+ ],
+ "score": 1.0,
+ "content": "’s ground-truth",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 520,
+ 504,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 504,
+ 532
+ ],
+ "score": 1.0,
+ "content": "MLP parameters randomly while leaving other variables’ MLPs untouched. A copy of the old",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 543
+ ],
+ "score": 1.0,
+ "content": "parameters is saved, allowing the intervention to be undone by resetting the parameters back to their",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 542,
+ 170,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 170,
+ 554
+ ],
+ "score": 1.0,
+ "content": "original values.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 565,
+ 505,
+ 643
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 504,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 320,
+ 579
+ ],
+ "score": 1.0,
+ "content": "Predicting interventions After an intervention on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 566,
+ 333,
+ 577
+ ],
+ "score": 0.88,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 564,
+ 494,
+ 579
+ ],
+ "score": 1.0,
+ "content": ", the gradients into the learned model’s",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 494,
+ 568,
+ 504,
+ 577
+ ],
+ "score": 0.81,
+ "content": "\\gamma _ { i }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 222,
+ 588
+ ],
+ "score": 1.0,
+ "content": "and the slow weights for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 578,
+ 227,
+ 587
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 577,
+ 484,
+ 588
+ ],
+ "score": 1.0,
+ "content": "-th conditional are false, because they do not bear the blame for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 484,
+ 577,
+ 497,
+ 588
+ ],
+ "score": 0.89,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 497,
+ 577,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": "’s",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "outcome (which lies with the intervener). We find that ignoring this issue considerably hurts or slows",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 506,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 506,
+ 611
+ ],
+ "score": 1.0,
+ "content": "down meta-learning, suggesting that we should try to infer on which variable the intervention took",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 610,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "place. For this purpose, we take advantage of the fact that the conditional likelihood of the intervened",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 620,
+ 324,
+ 633
+ ],
+ "score": 1.0,
+ "content": "variable tends to have a poorer relative likelihood under",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 621,
+ 343,
+ 631
+ ],
+ "score": 0.9,
+ "content": "D _ { \\mathrm { i n t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 620,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": ", so we pick the variable with the greatest",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 267,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 267,
+ 644
+ ],
+ "score": 1.0,
+ "content": "deterioration in likelihood as our guess.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 656,
+ 230,
+ 667
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 232,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 232,
+ 669
+ ],
+ "score": 1.0,
+ "content": "3.6 TRAINING ALGORITHM",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 676,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 231,
+ 690
+ ],
+ "score": 1.0,
+ "content": "The structural meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 677,
+ 258,
+ 689
+ ],
+ "score": 0.92,
+ "content": "\\sigma ( \\gamma _ { i j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 676,
+ 444,
+ 690
+ ],
+ "score": 1.0,
+ "content": "represent the belief in the hypothesis that node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 678,
+ 449,
+ 687
+ ],
+ "score": 0.72,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 676,
+ 487,
+ 690
+ ],
+ "score": 1.0,
+ "content": "has node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 487,
+ 678,
+ 493,
+ 689
+ ],
+ "score": 0.82,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "as",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "a direct causal parent. We may sample from this belief, obtaining different configurations (causal",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "structures) of the causal graph. We hypothesize that the correct configuration enables better adaptation",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "to a slight change in distribution, e.g. resulting from a soft intervention. Hence, we evaluate different",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 470,
+ 734
+ ],
+ "score": 1.0,
+ "content": "configurations under the transfer distribution; those giving a higher transfer likelihood under",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 470,
+ 721,
+ 489,
+ 732
+ ],
+ "score": 0.9,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 490,
+ 719,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "get",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 39
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 39
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 463,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 465,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 465,
+ 97
+ ],
+ "score": 1.0,
+ "content": "We adopt for this purpose the gradient estimate of Bengio et al. (2019) (but see footnote 1):",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0,
+ "bbox_fs": [
+ 105,
+ 80,
+ 465,
+ 97
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 233,
+ 97,
+ 377,
+ 130
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 233,
+ 97,
+ 377,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 233,
+ 97,
+ 377,
+ 130
+ ],
+ "score": 0.95,
+ "content": "g _ { i j } = \\frac { \\sum _ { k } ( \\sigma ( \\gamma _ { i j } ) - c _ { i j } ^ { ( k ) } ) \\mathcal { L } _ { C , i } ^ { ( k ) } ( X ) } { \\sum _ { k } \\mathcal { L } _ { C , i } ^ { ( k ) } ( X ) }",
+ "type": "interline_equation",
+ "image_path": "05663c44b32bacf4e94ac7b3c8468e7e7e7174553d7cf2a1c4855974bd17f26a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 233,
+ 97,
+ 377,
+ 113.5
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 233,
+ 113.5,
+ 377,
+ 130.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 135,
+ 503,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 150,
+ 149
+ ],
+ "score": 1.0,
+ "content": "where the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 134,
+ 162,
+ 145
+ ],
+ "score": 0.53,
+ "content": "( k )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 132,
+ 366,
+ 149
+ ],
+ "score": 1.0,
+ "content": "superscript indicates the values obtained for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 136,
+ 373,
+ 145
+ ],
+ "score": 0.83,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 132,
+ 421,
+ 149
+ ],
+ "score": 1.0,
+ "content": "-th draw of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 422,
+ 136,
+ 430,
+ 145
+ ],
+ "score": 0.82,
+ "content": "C",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 431,
+ 132,
+ 506,
+ 149
+ ],
+ "score": 1.0,
+ "content": ". This gradient is",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 145,
+ 455,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 194,
+ 159
+ ],
+ "score": 1.0,
+ "content": "estimated solely with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 147,
+ 213,
+ 157
+ ],
+ "score": 0.91,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 145,
+ 333,
+ 159
+ ],
+ "score": 1.0,
+ "content": "because estimates employing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 333,
+ 147,
+ 339,
+ 156
+ ],
+ "score": 0.79,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 145,
+ 455,
+ 159
+ ],
+ "score": 1.0,
+ "content": "have much greater variance.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 132,
+ 506,
+ 159
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 189
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 166,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "Acyclic Constraint We add a regularization term to the loss term that discourages the model from",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 423,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 423,
+ 191
+ ],
+ "score": 1.0,
+ "content": "having length-2 cycles in the learned adjacency matrix. The regularizer term is",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 166,
+ 505,
+ 191
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 186,
+ 193,
+ 425,
+ 221
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 186,
+ 193,
+ 425,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 186,
+ 193,
+ 425,
+ 221
+ ],
+ "score": 0.94,
+ "content": "J _ { \\mathrm { D A G } } = \\sum _ { i \\neq j } \\cosh ( \\sigma ( \\gamma _ { i j } ) \\sigma ( \\gamma _ { j i } ) ) , \\quad \\forall i , j \\in \\{ 0 , \\ldots , M - 1 \\}",
+ "type": "interline_equation",
+ "image_path": "c9e6c645bf54bf607c6162b9ac69233dcfc2d763372ff5553590f369bc6a63c0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 186,
+ 193,
+ 425,
+ 207.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 186,
+ 207.0,
+ 425,
+ 221.0
+ ],
+ "spans": [],
+ "index": 8
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 225,
+ 465,
+ 237
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 466,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 466,
+ 238
+ ],
+ "score": 1.0,
+ "content": "and is derived from Zheng et al. (2018). The details of the derivation are in the Appendix.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 105,
+ 223,
+ 466,
+ 238
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 250,
+ 226,
+ 261
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 249,
+ 227,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 249,
+ 227,
+ 262
+ ],
+ "score": 1.0,
+ "content": "3.4 MODEL DESCRIPTION",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 270,
+ 506,
+ 370
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 241,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Learner Model We model the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 271,
+ 253,
+ 280
+ ],
+ "score": 0.78,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 270,
+ 348,
+ 284
+ ],
+ "score": 1.0,
+ "content": "structural assignments",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 349,
+ 270,
+ 446,
+ 283
+ ],
+ "score": 0.92,
+ "content": "X _ { i } : = f _ { i } ( X _ { p a ( i , C ) } , N _ { i } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 446,
+ 270,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "(Eq. 1) of the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 281,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 281,
+ 169,
+ 293
+ ],
+ "score": 1.0,
+ "content": "SCM as a set of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 282,
+ 181,
+ 291
+ ],
+ "score": 0.79,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 281,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "multi-layer perceptrons (MLPs), as in Bengio et al. (2019). The MLPs are identical",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 292,
+ 507,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 292,
+ 507,
+ 305
+ ],
+ "score": 1.0,
+ "content": "in shape but do not share any parameters, since they are modeling independent causal mechanisms.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 236,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Each possesses an input layer of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 303,
+ 269,
+ 313
+ ],
+ "score": 0.91,
+ "content": "M \\times N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 269,
+ 303,
+ 321,
+ 316
+ ],
+ "score": 1.0,
+ "content": "neurons (for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 304,
+ 333,
+ 313
+ ],
+ "score": 0.82,
+ "content": "M",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 303,
+ 434,
+ 316
+ ],
+ "score": 1.0,
+ "content": "one-hot vectors of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 435,
+ 304,
+ 445,
+ 313
+ ],
+ "score": 0.79,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 445,
+ 303,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "each), a single",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 506,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 270,
+ 327
+ ],
+ "score": 1.0,
+ "content": "hidden layer chosen arbitrarily to have",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 270,
+ 314,
+ 331,
+ 326
+ ],
+ "score": 0.9,
+ "content": "\\operatorname* { m a x } ( 4 M , 4 N )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 313,
+ 506,
+ 327
+ ],
+ "score": 1.0,
+ "content": "neurons with a LeakyReLU activation of",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 325,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 237,
+ 338
+ ],
+ "score": 1.0,
+ "content": "slope 0.1, and an output layer of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 326,
+ 248,
+ 335
+ ],
+ "score": 0.84,
+ "content": "N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 325,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "neurons representing the unnormalized log-probabilities of each",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 504,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 180,
+ 349
+ ],
+ "score": 1.0,
+ "content": "category. To force",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 336,
+ 189,
+ 348
+ ],
+ "score": 0.87,
+ "content": "f _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 335,
+ 364,
+ 349
+ ],
+ "score": 1.0,
+ "content": "to rely exclusively on the direct ancestor set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 336,
+ 399,
+ 348
+ ],
+ "score": 0.92,
+ "content": "p a ( i , C )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 335,
+ 495,
+ 349
+ ],
+ "score": 1.0,
+ "content": "under adjacency matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 336,
+ 504,
+ 346
+ ],
+ "score": 0.79,
+ "content": "C",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 347,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 347,
+ 258,
+ 359
+ ],
+ "score": 1.0,
+ "content": "(See Eqn. 2), the one-hot input vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 348,
+ 272,
+ 359
+ ],
+ "score": 0.89,
+ "content": "X _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 347,
+ 322,
+ 359
+ ],
+ "score": 1.0,
+ "content": "for variable",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 347,
+ 335,
+ 358
+ ],
+ "score": 0.87,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 347,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": "’s MLP is masked by the Boolean element",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 357,
+ 462,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 119,
+ 370
+ ],
+ "score": 0.86,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 357,
+ 322,
+ 372
+ ],
+ "score": 1.0,
+ "content": ". The functional parameters of the MLP are the set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 358,
+ 459,
+ 370
+ ],
+ "score": 0.89,
+ "content": "\\theta = \\left\\{ { \\cal W } 0 _ { i h j n } , { \\tt B } 0 _ { i h } , { \\tt W } 1 _ { i n h } , { \\tt B } 1 _ { i n } \\right\\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 459,
+ 357,
+ 462,
+ 372
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 270,
+ 507,
+ 372
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 374,
+ 504,
+ 397
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 298,
+ 387
+ ],
+ "score": 1.0,
+ "content": "An example of the multi-MLP architecture with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 298,
+ 375,
+ 328,
+ 385
+ ],
+ "score": 0.9,
+ "content": "M = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 374,
+ 424,
+ 387
+ ],
+ "score": 1.0,
+ "content": "categorical variables of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 375,
+ 452,
+ 385
+ ],
+ "score": 0.9,
+ "content": "N = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 374,
+ 506,
+ 387
+ ],
+ "score": 1.0,
+ "content": "categories is",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 386,
+ 184,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 386,
+ 184,
+ 398
+ ],
+ "score": 1.0,
+ "content": "shown in Figure 1.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 106,
+ 374,
+ 506,
+ 398
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 409,
+ 505,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 409,
+ 506,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 409,
+ 506,
+ 422
+ ],
+ "score": 1.0,
+ "content": "Ground-Truth Model In our experiments, ground-truth SCM models are parametrized as a set of",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 420,
+ 505,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 420,
+ 505,
+ 432
+ ],
+ "score": 1.0,
+ "content": "MLPs of the same size as the learner models, thus avoid having to manually define the Conditional",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 444
+ ],
+ "score": 1.0,
+ "content": "Probability Tables (CPTs). Ground-truth models exist in two variants that differ mainly in initial-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 222,
+ 455
+ ],
+ "score": 1.0,
+ "content": "ization: synthetic, where the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 223,
+ 443,
+ 229,
+ 452
+ ],
+ "score": 0.77,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 442,
+ 360,
+ 455
+ ],
+ "score": 1.0,
+ "content": "are randomly initialized and the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 444,
+ 368,
+ 453
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 442,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "are either randomly-initialized or",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 452,
+ 506,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 273,
+ 466
+ ],
+ "score": 1.0,
+ "content": "pre-specified; and real-world, where the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 273,
+ 453,
+ 280,
+ 463
+ ],
+ "score": 0.78,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 280,
+ 452,
+ 298,
+ 466
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 455,
+ 306,
+ 465
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 452,
+ 506,
+ 466
+ ],
+ "score": 1.0,
+ "content": "are both initialized so as to closely replicate the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 465,
+ 246,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 246,
+ 476
+ ],
+ "score": 1.0,
+ "content": "CPTs of given Bayesian networks.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 409,
+ 506,
+ 476
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 488,
+ 202,
+ 500
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 203,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 203,
+ 501
+ ],
+ "score": 1.0,
+ "content": "3.5 INTERVENTIONS",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 509,
+ 505,
+ 554
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 504,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 352,
+ 522
+ ],
+ "score": 1.0,
+ "content": "Soft interventions To execute an intervention on variable",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 510,
+ 364,
+ 520
+ ],
+ "score": 0.88,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 509,
+ 430,
+ 522
+ ],
+ "score": 1.0,
+ "content": ", we reinitialize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 510,
+ 442,
+ 520
+ ],
+ "score": 0.87,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 443,
+ 509,
+ 504,
+ 522
+ ],
+ "score": 1.0,
+ "content": "’s ground-truth",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 520,
+ 504,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 504,
+ 532
+ ],
+ "score": 1.0,
+ "content": "MLP parameters randomly while leaving other variables’ MLPs untouched. A copy of the old",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 506,
+ 543
+ ],
+ "score": 1.0,
+ "content": "parameters is saved, allowing the intervention to be undone by resetting the parameters back to their",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 542,
+ 170,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 170,
+ 554
+ ],
+ "score": 1.0,
+ "content": "original values.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 509,
+ 506,
+ 554
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 565,
+ 505,
+ 643
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 504,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 320,
+ 579
+ ],
+ "score": 1.0,
+ "content": "Predicting interventions After an intervention on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 566,
+ 333,
+ 577
+ ],
+ "score": 0.88,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 333,
+ 564,
+ 494,
+ 579
+ ],
+ "score": 1.0,
+ "content": ", the gradients into the learned model’s",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 494,
+ 568,
+ 504,
+ 577
+ ],
+ "score": 0.81,
+ "content": "\\gamma _ { i }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 577,
+ 505,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 577,
+ 222,
+ 588
+ ],
+ "score": 1.0,
+ "content": "and the slow weights for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 578,
+ 227,
+ 587
+ ],
+ "score": 0.66,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 577,
+ 484,
+ 588
+ ],
+ "score": 1.0,
+ "content": "-th conditional are false, because they do not bear the blame for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 484,
+ 577,
+ 497,
+ 588
+ ],
+ "score": 0.89,
+ "content": "X _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 497,
+ 577,
+ 505,
+ 588
+ ],
+ "score": 1.0,
+ "content": "’s",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "outcome (which lies with the intervener). We find that ignoring this issue considerably hurts or slows",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 506,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 506,
+ 611
+ ],
+ "score": 1.0,
+ "content": "down meta-learning, suggesting that we should try to infer on which variable the intervention took",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 610,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "place. For this purpose, we take advantage of the fact that the conditional likelihood of the intervened",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 620,
+ 324,
+ 633
+ ],
+ "score": 1.0,
+ "content": "variable tends to have a poorer relative likelihood under",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 324,
+ 621,
+ 343,
+ 631
+ ],
+ "score": 0.9,
+ "content": "D _ { \\mathrm { i n t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 344,
+ 620,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": ", so we pick the variable with the greatest",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 267,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 267,
+ 644
+ ],
+ "score": 1.0,
+ "content": "deterioration in likelihood as our guess.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 564,
+ 506,
+ 644
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 656,
+ 230,
+ 667
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 232,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 232,
+ 669
+ ],
+ "score": 1.0,
+ "content": "3.6 TRAINING ALGORITHM",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 676,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 231,
+ 690
+ ],
+ "score": 1.0,
+ "content": "The structural meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 677,
+ 258,
+ 689
+ ],
+ "score": 0.92,
+ "content": "\\sigma ( \\gamma _ { i j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 676,
+ 444,
+ 690
+ ],
+ "score": 1.0,
+ "content": "represent the belief in the hypothesis that node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 678,
+ 449,
+ 687
+ ],
+ "score": 0.72,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 450,
+ 676,
+ 487,
+ 690
+ ],
+ "score": 1.0,
+ "content": "has node",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 487,
+ 678,
+ 493,
+ 689
+ ],
+ "score": 0.82,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 676,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "as",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "a direct causal parent. We may sample from this belief, obtaining different configurations (causal",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "structures) of the causal graph. We hypothesize that the correct configuration enables better adaptation",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "to a slight change in distribution, e.g. resulting from a soft intervention. Hence, we evaluate different",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 470,
+ 734
+ ],
+ "score": 1.0,
+ "content": "configurations under the transfer distribution; those giving a higher transfer likelihood under",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 470,
+ 721,
+ 489,
+ 732
+ ],
+ "score": 0.9,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 490,
+ 719,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "get",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "a higher reward and their probability is increased. The functional (meta-)parameters are trained as",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "usual by gradient ascent on the log-likelihood. The details of the training algorithm is in Section A.2",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 104,
+ 176,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 104,
+ 176,
+ 117
+ ],
+ "score": 1.0,
+ "content": "in the Appendix.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 105,
+ 676,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "a higher reward and their probability is increased. The functional (meta-)parameters are trained as",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "usual by gradient ascent on the log-likelihood. The details of the training algorithm is in Section A.2",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 104,
+ 176,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 104,
+ 176,
+ 117
+ ],
+ "score": 1.0,
+ "content": "in the Appendix.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 121,
+ 505,
+ 187
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 506,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 342,
+ 134
+ ],
+ "score": 1.0,
+ "content": "Synthetic datasets use either a specified edge structure for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 124,
+ 349,
+ 133
+ ],
+ "score": 0.8,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 350,
+ 121,
+ 443,
+ 134
+ ],
+ "score": 1.0,
+ "content": ", or randomly initialize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 123,
+ 457,
+ 134
+ ],
+ "score": 0.87,
+ "content": "\\gamma _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 121,
+ 506,
+ 134
+ ],
+ "score": 1.0,
+ "content": "such that it",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "is Boolean, strictly lower-triangular, and each row has an expected sum of 1-5 (and therefore each",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 155
+ ],
+ "score": 1.0,
+ "content": "node expects 1-5 direct ancestors). Real-world datasets, specified as CPTs, must first be converted or",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 502,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 495,
+ 168
+ ],
+ "score": 1.0,
+ "content": "approximated by a near-ground-truth MLP. We use the graph’s proper edge structure to initialize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 156,
+ 502,
+ 166
+ ],
+ "score": 0.77,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 165,
+ 504,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 165,
+ 145,
+ 177
+ ],
+ "score": 1.0,
+ "content": "and learn",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 166,
+ 151,
+ 175
+ ],
+ "score": 0.73,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 165,
+ 504,
+ 177
+ ],
+ "score": 1.0,
+ "content": "by training individually each MLP in the set to replicate the correct pre-softmax logits. In",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 353,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 353,
+ 188
+ ],
+ "score": 1.0,
+ "content": "practice, excellent reproductions of the CPT can be achieved.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 193,
+ 505,
+ 281
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 193,
+ 505,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 193,
+ 505,
+ 205
+ ],
+ "score": 1.0,
+ "content": "Stability of training Our model requires simultaneous training of both the structural and the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 216
+ ],
+ "score": 1.0,
+ "content": "functional meta-parameters, but these are not independent and do influence each other, which leads",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 271,
+ 228
+ ],
+ "score": 1.0,
+ "content": "to instability in training. For example, if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 271,
+ 215,
+ 316,
+ 227
+ ],
+ "score": 0.93,
+ "content": "\\sigma ( \\gamma _ { i j } ) \\approx 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 215,
+ 379,
+ 228
+ ],
+ "score": 1.0,
+ "content": "incorrectly, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 216,
+ 384,
+ 225
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 215,
+ 506,
+ 228
+ ],
+ "score": 1.0,
+ "content": "-th MLP does not learn to use",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 129,
+ 239
+ ],
+ "score": 1.0,
+ "content": "input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 130,
+ 226,
+ 143,
+ 238
+ ],
+ "score": 0.88,
+ "content": "X _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 226,
+ 234,
+ 239
+ ],
+ "score": 1.0,
+ "content": ", and vice-versa, if the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 227,
+ 239,
+ 236
+ ],
+ "score": 0.76,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 239,
+ 226,
+ 424,
+ 239
+ ],
+ "score": 1.0,
+ "content": "-th MLP has not learned to properly use input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 226,
+ 437,
+ 238
+ ],
+ "score": 0.89,
+ "content": "X _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 226,
+ 505,
+ 239
+ ],
+ "score": 1.0,
+ "content": ", this will favour",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 506,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 141,
+ 249
+ ],
+ "score": 1.0,
+ "content": "pushing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 141,
+ 237,
+ 168,
+ 249
+ ],
+ "score": 0.92,
+ "content": "\\sigma ( \\gamma _ { i j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 237,
+ 506,
+ 249
+ ],
+ "score": 1.0,
+ "content": "towards 0. To overcome this instability, we pretrain the model under observational",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 260
+ ],
+ "score": 1.0,
+ "content": "data (from the distribution of the data before interventions) using dropout on the inputs. This ensures",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 255,
+ 271
+ ],
+ "score": 1.0,
+ "content": "that the functional meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 255,
+ 259,
+ 274,
+ 270
+ ],
+ "score": 0.91,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 258,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": "are not too biased towards certain configurations of the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 187,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 175,
+ 284
+ ],
+ "score": 1.0,
+ "content": "meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 272,
+ 182,
+ 281
+ ],
+ "score": 0.79,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 270,
+ 187,
+ 284
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 290,
+ 211,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 213,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 213,
+ 305
+ ],
+ "score": 1.0,
+ "content": "4 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 308,
+ 506,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 321
+ ],
+ "score": 1.0,
+ "content": "The recovery of the underlying structural causal graph from observational and interventional data is",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "a fundamental problem (Pearl, 1995; 2009). Different approaches have been studied, score-based,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 331,
+ 506,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 331,
+ 506,
+ 342
+ ],
+ "score": 1.0,
+ "content": "constraint-based and asymmetry-based methods. Score-based methods search through the space of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 341,
+ 506,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 341,
+ 506,
+ 353
+ ],
+ "score": 1.0,
+ "content": "all possible directed acyclic graphs (DAGs) representing the causal structure based on some form of",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "scoring function for network structures (Chickering, 2002; Tsamardinos et al., 2006; Goudet et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 376
+ ],
+ "score": 1.0,
+ "content": "2017; Hauser & Bühlmann, 2012; Heckerman et al., 1995; Cooper & Yoo, 1999; Hauser & Bühlmann,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "2012). Constraint-based methods (Spirtes et al., 2000; Sun et al., 2007; Zhang et al., 2012) infer the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 398
+ ],
+ "score": 1.0,
+ "content": "DAG by analyzing the conditional independence of data. Eaton & Murphy (2007b) use dynamic",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "programming techniques to accelerate Markov Chain Monte Carlo (MCMC) sampling in a Bayesian",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "approach to structure learning for discrete variable DAGs. Asymmetry-based methods (Shimizu et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "2006; Hoyer et al., 2009; Daniusis et al., 2012; Budhathoki & Vreeken, 2017; Mitrovic et al., 2018)",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "assume asymmetry between cause and effect in the data and try to use this information to estimate the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "causal structure. Recently Peters et al. (2016); Ghassami et al. (2017) proposed to exploit invariance",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 463
+ ],
+ "score": 1.0,
+ "content": "across different environments to infer causal structure, but are difficult to scale to large graphs due to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 462,
+ 396,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 462,
+ 396,
+ 474
+ ],
+ "score": 1.0,
+ "content": "the necessary iteration over the super-exponential set of possible graphs.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 478,
+ 505,
+ 533
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 491
+ ],
+ "score": 1.0,
+ "content": "For interventional data, it is often assumed that the models have access to full intervention information,",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 501
+ ],
+ "score": 1.0,
+ "content": "which is rare in the real world. Rothenhäusler et al. (2015) have investigated the case of additive",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 499,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 499,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "shift interventions, while Eaton & Murphy (2007a) have examined the situation where the targets of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "experimental interventions are imperfect or uncertain. This is different from our setting where the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 522,
+ 506,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 522,
+ 506,
+ 535
+ ],
+ "score": 1.0,
+ "content": "intervention is unknown to start with and is assumed to arise from other agents and the environment.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 539,
+ 505,
+ 627
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "Learning based methods have been proposed (Guyon, a;b; Lopez-Paz et al., 2015) and there also exist",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "recent approaches using the generalization ability of neural networks to learn causal signals from",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "purely observational data (Kalainathan et al., 2018; Goudet et al., 2018). Neural network methods",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "equipped with learned masks, such as (Ivanov et al., 2018; Li et al., 2019; Yoon et al., 2018; Douglas",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "et al., 2017), exist in the literature, but only a few (Kalainathan et al., 2018) have been adapted to",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "causal inference. This last work is, however, tailored for causal inference on continuous variables",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "and from observations only. Adapting it to a discrete-variable setting is made difficult by its use of a",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 616,
+ 417,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 616,
+ 417,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Generative Adversarial Network (GAN) Goodfellow et al. (2014) framework.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 41.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 632,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "Turning now to meta-learning, Dasgupta et al. (2019) have used it to learn to make predictions under",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "interventions. However, their approach does not induce a causal graph, neither explicitly nor via",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "decoding. Thus, it cannot be used for general causal discovery, but only to make predictions of variable",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "values. Most similar to our work, Bengio et al. (2019) proposes a meta-learning framework for",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "learning causal models from interventional data. However, the proposed method (Bengio et al., 2019)",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "explicitly models every possible set of parents for every child variable and attempts to distinguish the",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "best among them. Because there are combinatorially-many such parent sets, the method cannot scale",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "beyond trivial graphs. In our work, we bypass this restriction by modeling the edge between any 2",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "variables as a dropout probability and hence our model only scales quadratically with the graph size.",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [],
+ "index": 1,
+ "bbox_fs": [
+ 106,
+ 82,
+ 505,
+ 117
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 121,
+ 505,
+ 187
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 506,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 342,
+ 134
+ ],
+ "score": 1.0,
+ "content": "Synthetic datasets use either a specified edge structure for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 124,
+ 349,
+ 133
+ ],
+ "score": 0.8,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 350,
+ 121,
+ 443,
+ 134
+ ],
+ "score": 1.0,
+ "content": ", or randomly initialize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 123,
+ 457,
+ 134
+ ],
+ "score": 0.87,
+ "content": "\\gamma _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 121,
+ 506,
+ 134
+ ],
+ "score": 1.0,
+ "content": "such that it",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "is Boolean, strictly lower-triangular, and each row has an expected sum of 1-5 (and therefore each",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 505,
+ 155
+ ],
+ "score": 1.0,
+ "content": "node expects 1-5 direct ancestors). Real-world datasets, specified as CPTs, must first be converted or",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 502,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 495,
+ 168
+ ],
+ "score": 1.0,
+ "content": "approximated by a near-ground-truth MLP. We use the graph’s proper edge structure to initialize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 156,
+ 502,
+ 166
+ ],
+ "score": 0.77,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 165,
+ 504,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 165,
+ 145,
+ 177
+ ],
+ "score": 1.0,
+ "content": "and learn",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 166,
+ 151,
+ 175
+ ],
+ "score": 0.73,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 165,
+ 504,
+ 177
+ ],
+ "score": 1.0,
+ "content": "by training individually each MLP in the set to replicate the correct pre-softmax logits. In",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 176,
+ 353,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 353,
+ 188
+ ],
+ "score": 1.0,
+ "content": "practice, excellent reproductions of the CPT can be achieved.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 121,
+ 506,
+ 188
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 193,
+ 505,
+ 281
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 193,
+ 505,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 193,
+ 505,
+ 205
+ ],
+ "score": 1.0,
+ "content": "Stability of training Our model requires simultaneous training of both the structural and the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 204,
+ 505,
+ 216
+ ],
+ "score": 1.0,
+ "content": "functional meta-parameters, but these are not independent and do influence each other, which leads",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 271,
+ 228
+ ],
+ "score": 1.0,
+ "content": "to instability in training. For example, if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 271,
+ 215,
+ 316,
+ 227
+ ],
+ "score": 0.93,
+ "content": "\\sigma ( \\gamma _ { i j } ) \\approx 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 215,
+ 379,
+ 228
+ ],
+ "score": 1.0,
+ "content": "incorrectly, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 216,
+ 384,
+ 225
+ ],
+ "score": 0.71,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 215,
+ 506,
+ 228
+ ],
+ "score": 1.0,
+ "content": "-th MLP does not learn to use",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 129,
+ 239
+ ],
+ "score": 1.0,
+ "content": "input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 130,
+ 226,
+ 143,
+ 238
+ ],
+ "score": 0.88,
+ "content": "X _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 144,
+ 226,
+ 234,
+ 239
+ ],
+ "score": 1.0,
+ "content": ", and vice-versa, if the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 227,
+ 239,
+ 236
+ ],
+ "score": 0.76,
+ "content": "i",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 239,
+ 226,
+ 424,
+ 239
+ ],
+ "score": 1.0,
+ "content": "-th MLP has not learned to properly use input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 226,
+ 437,
+ 238
+ ],
+ "score": 0.89,
+ "content": "X _ { j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 226,
+ 505,
+ 239
+ ],
+ "score": 1.0,
+ "content": ", this will favour",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 237,
+ 506,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 141,
+ 249
+ ],
+ "score": 1.0,
+ "content": "pushing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 141,
+ 237,
+ 168,
+ 249
+ ],
+ "score": 0.92,
+ "content": "\\sigma ( \\gamma _ { i j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 237,
+ 506,
+ 249
+ ],
+ "score": 1.0,
+ "content": "towards 0. To overcome this instability, we pretrain the model under observational",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 260
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 260
+ ],
+ "score": 1.0,
+ "content": "data (from the distribution of the data before interventions) using dropout on the inputs. This ensures",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 255,
+ 271
+ ],
+ "score": 1.0,
+ "content": "that the functional meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 255,
+ 259,
+ 274,
+ 270
+ ],
+ "score": 0.91,
+ "content": "\\theta _ { \\mathrm { s l o w } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 258,
+ 505,
+ 271
+ ],
+ "score": 1.0,
+ "content": "are not too biased towards certain configurations of the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 270,
+ 187,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 175,
+ 284
+ ],
+ "score": 1.0,
+ "content": "meta-parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 272,
+ 182,
+ 281
+ ],
+ "score": 0.79,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 270,
+ 187,
+ 284
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 105,
+ 193,
+ 506,
+ 284
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 290,
+ 211,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 213,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 213,
+ 305
+ ],
+ "score": 1.0,
+ "content": "4 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 308,
+ 506,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 321
+ ],
+ "score": 1.0,
+ "content": "The recovery of the underlying structural causal graph from observational and interventional data is",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "a fundamental problem (Pearl, 1995; 2009). Different approaches have been studied, score-based,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 331,
+ 506,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 331,
+ 506,
+ 342
+ ],
+ "score": 1.0,
+ "content": "constraint-based and asymmetry-based methods. Score-based methods search through the space of",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 341,
+ 506,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 341,
+ 506,
+ 353
+ ],
+ "score": 1.0,
+ "content": "all possible directed acyclic graphs (DAGs) representing the causal structure based on some form of",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "scoring function for network structures (Chickering, 2002; Tsamardinos et al., 2006; Goudet et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 376
+ ],
+ "score": 1.0,
+ "content": "2017; Hauser & Bühlmann, 2012; Heckerman et al., 1995; Cooper & Yoo, 1999; Hauser & Bühlmann,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "2012). Constraint-based methods (Spirtes et al., 2000; Sun et al., 2007; Zhang et al., 2012) infer the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 398
+ ],
+ "score": 1.0,
+ "content": "DAG by analyzing the conditional independence of data. Eaton & Murphy (2007b) use dynamic",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "programming techniques to accelerate Markov Chain Monte Carlo (MCMC) sampling in a Bayesian",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "approach to structure learning for discrete variable DAGs. Asymmetry-based methods (Shimizu et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "2006; Hoyer et al., 2009; Daniusis et al., 2012; Budhathoki & Vreeken, 2017; Mitrovic et al., 2018)",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 506,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 506,
+ 441
+ ],
+ "score": 1.0,
+ "content": "assume asymmetry between cause and effect in the data and try to use this information to estimate the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "causal structure. Recently Peters et al. (2016); Ghassami et al. (2017) proposed to exploit invariance",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 463
+ ],
+ "score": 1.0,
+ "content": "across different environments to infer causal structure, but are difficult to scale to large graphs due to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 462,
+ 396,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 462,
+ 396,
+ 474
+ ],
+ "score": 1.0,
+ "content": "the necessary iteration over the super-exponential set of possible graphs.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 105,
+ 307,
+ 506,
+ 474
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 478,
+ 505,
+ 533
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 491
+ ],
+ "score": 1.0,
+ "content": "For interventional data, it is often assumed that the models have access to full intervention information,",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 501
+ ],
+ "score": 1.0,
+ "content": "which is rare in the real world. Rothenhäusler et al. (2015) have investigated the case of additive",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 499,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 499,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "shift interventions, while Eaton & Murphy (2007a) have examined the situation where the targets of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "experimental interventions are imperfect or uncertain. This is different from our setting where the",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 522,
+ 506,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 522,
+ 506,
+ 535
+ ],
+ "score": 1.0,
+ "content": "intervention is unknown to start with and is assumed to arise from other agents and the environment.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 478,
+ 506,
+ 535
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 539,
+ 505,
+ 627
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "Learning based methods have been proposed (Guyon, a;b; Lopez-Paz et al., 2015) and there also exist",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "recent approaches using the generalization ability of neural networks to learn causal signals from",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 574
+ ],
+ "score": 1.0,
+ "content": "purely observational data (Kalainathan et al., 2018; Goudet et al., 2018). Neural network methods",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "equipped with learned masks, such as (Ivanov et al., 2018; Li et al., 2019; Yoon et al., 2018; Douglas",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "et al., 2017), exist in the literature, but only a few (Kalainathan et al., 2018) have been adapted to",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "causal inference. This last work is, however, tailored for causal inference on continuous variables",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "and from observations only. Adapting it to a discrete-variable setting is made difficult by its use of a",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 616,
+ 417,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 616,
+ 417,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Generative Adversarial Network (GAN) Goodfellow et al. (2014) framework.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 41.5,
+ "bbox_fs": [
+ 105,
+ 540,
+ 506,
+ 628
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 632,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "Turning now to meta-learning, Dasgupta et al. (2019) have used it to learn to make predictions under",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "interventions. However, their approach does not induce a causal graph, neither explicitly nor via",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "decoding. Thus, it cannot be used for general causal discovery, but only to make predictions of variable",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "values. Most similar to our work, Bengio et al. (2019) proposes a meta-learning framework for",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "learning causal models from interventional data. However, the proposed method (Bengio et al., 2019)",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "explicitly models every possible set of parents for every child variable and attempts to distinguish the",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "best among them. Because there are combinatorially-many such parent sets, the method cannot scale",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "beyond trivial graphs. In our work, we bypass this restriction by modeling the edge between any 2",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "variables as a dropout probability and hence our model only scales quadratically with the graph size.",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ }
+ ],
+ "index": 50,
+ "bbox_fs": [
+ 105,
+ 633,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 82,
+ 316,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 317,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 317,
+ 95
+ ],
+ "score": 1.0,
+ "content": "5 EXPERIMENTAL SETUP AND RESULTS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 112
+ ],
+ "score": 1.0,
+ "content": "Our experiments aim to evaluate the proposed method to recover the correct causal structure and",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "which elements of the method matter. We first evaluate our model on a synthetic dataset where we",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "have control over the number of variables and causal edges in the ground-truth SCM. This allows",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 473,
+ 146
+ ],
+ "score": 1.0,
+ "content": "us to verify after learning with what accuracy we recover the individual decisions about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 134,
+ 486,
+ 145
+ ],
+ "score": 0.87,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 132,
+ 506,
+ 146
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "understand the performance of our algorithm under various conditions. We then evaluate our method",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 154,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 154,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": "on real world datasets collected from the BnLearn dataset repository, and show that the proposed",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "approach recovers the true causal structure. We then show that the trained models can correctly",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "predict the consequences of previously unseen interventions on the rest of the graph. We also perform",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 188,
+ 373,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 188,
+ 373,
+ 200
+ ],
+ "score": 1.0,
+ "content": "ablations showing how important is each component of the model.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 212,
+ 228,
+ 222
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 211,
+ 229,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 211,
+ 229,
+ 224
+ ],
+ "score": 1.0,
+ "content": "5.1 SYNTHETIC DATASETS",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 267
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 267
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 267
+ ],
+ "score": 0.964,
+ "type": "image",
+ "image_path": "f99f187fc376946f89ba5b065c406dc52459b277e1a3d71f66784cd8ed754b6a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 239.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 161,
+ 239.0,
+ 447,
+ 253.0
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 161,
+ 253.0,
+ 447,
+ 267.0
+ ],
+ "spans": [],
+ "index": 13
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 275,
+ 504,
+ 293
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Learned edges at three different stages of training. Left: Chain graph with 4 variables. Right: Fully-connected DAG graph with 4",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 283,
+ 136,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 283,
+ 136,
+ 292
+ ],
+ "score": 1.0,
+ "content": "variables.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14.5
+ }
+ ],
+ "index": 13.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 296,
+ 505,
+ 395
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "We first evaluate the model’s performance on several randomly-initialized SCMs with specific,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 252,
+ 320
+ ],
+ "score": 1.0,
+ "content": "representative graph structures. For",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 253,
+ 307,
+ 283,
+ 317
+ ],
+ "score": 0.91,
+ "content": "M = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 306,
+ 506,
+ 320
+ ],
+ "score": 1.0,
+ "content": "-variable DAGs, we consider every possible connected",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "graph: chain3, fork3, collider3 and confounder3 (See Fig. 7 in appendix). They exhibit",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "every graph sub-structure that can exist in larger graphs, and must be mastered before tackling larger",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 340,
+ 506,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 506,
+ 353
+ ],
+ "score": 1.0,
+ "content": "graphs. Since the number of possible DAGs grows super-exponentially with the number of variables,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 351,
+ 506,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 120,
+ 363
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 121,
+ 351,
+ 151,
+ 361
+ ],
+ "score": 0.88,
+ "content": "M > 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 351,
+ 506,
+ 363
+ ],
+ "score": 1.0,
+ "content": "up to 8 a selection of representative and edge-case graphs are chosen. The chainM and",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 170,
+ 374
+ ],
+ "score": 1.0,
+ "content": "fullM graphs",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 362,
+ 204,
+ 372
+ ],
+ "score": 0.85,
+ "content": "( \\mathbb { M } = 3 - 8",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": ") are the minimally and maximally connected M-variable graphs, while the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "remaining graphs are randomly generated with a varying sparsity level (1-4 expected number of",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 402,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 402,
+ 396
+ ],
+ "score": 1.0,
+ "content": "parents per node). The details of the setup can be found in Appendix A.5.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 406,
+ 505,
+ 484
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 406,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 406,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "Results The model can successfully recover the correct edge structure for all synthetic graphs con-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "sidered. The learning curves plotting the average cross-entropy (CE) loss for the learned edges against",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 219,
+ 441
+ ],
+ "score": 1.0,
+ "content": "the ground-truth model for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 220,
+ 429,
+ 252,
+ 439
+ ],
+ "score": 0.9,
+ "content": "M = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 428,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "are shown in Figure 7. The fully-connected confounder3",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 452
+ ],
+ "score": 1.0,
+ "content": "graph is particularly easy to learn, but all 3-variable graphs are learned perfectly. Plots of the same",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "process on 4- through 8-variable graphs were similarly encouraging, with all models converging to",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 506,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 506,
+ 475
+ ],
+ "score": 1.0,
+ "content": "a negligible loss. The results are, however, sensitive to some hyperparameters, notably the DAG",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 237,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 237,
+ 486
+ ],
+ "score": 1.0,
+ "content": "penalty and the sparsity penalty.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 496,
+ 289,
+ 507
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 291,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 291,
+ 509
+ ],
+ "score": 1.0,
+ "content": "5.2 REAL-WORLD DATASETS: BNLEARN",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 517,
+ 357,
+ 594
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 357,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 357,
+ 529
+ ],
+ "score": 1.0,
+ "content": "The Bayesian Network Repository is a collection of commonly-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 528,
+ 357,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 357,
+ 539
+ ],
+ "score": 1.0,
+ "content": "used causal Bayesian networks from the literature, suitable for",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 539,
+ 357,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 539,
+ 357,
+ 551
+ ],
+ "score": 1.0,
+ "content": "Bayesian and causal learning benchmarks. We evaluate our",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 549,
+ 357,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 357,
+ 562
+ ],
+ "score": 1.0,
+ "content": "model on the Earthquake (Korb & Nicholson, 2010), Cancer",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 560,
+ 357,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 357,
+ 573
+ ],
+ "score": 1.0,
+ "content": "(Korb & Nicholson, 2010) and Asia (Lauritzen & Spiegelhalter,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 571,
+ 358,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 571,
+ 172,
+ 584
+ ],
+ "score": 1.0,
+ "content": "1988) datasets (",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 572,
+ 201,
+ 582
+ ],
+ "score": 0.78,
+ "content": "M = 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 571,
+ 358,
+ 584
+ ],
+ "score": 1.0,
+ "content": ", 5 and 8-variables respectively, max-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 358,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 358,
+ 597
+ ],
+ "score": 1.0,
+ "content": "imum 2 parents per node) in the BnLearn dataset repository.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 595,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "The ground-truth SCM is given for each dataset, and the functional parameters are represented as",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "conditional probability tables (CPTs). We learn a near-ground-truth MLP from the dataset’s CPT and",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 616,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 616,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "use it as the ground-truth data generator. We also insert a (greater than 1) temperature factor in order",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 104,
+ 625,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 625,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "to increase the likelihood of sampling some very rare events in the CPTs. Details of the setup can be",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 212,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 212,
+ 650
+ ],
+ "score": 1.0,
+ "content": "found in Appendix A.5.1.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 565
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 565
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 565
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "d9b5c523ea9b4196e4b0c1585be866e7e439433e299ccc6ecc2da33b588016ef.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 542.5
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 365,
+ 542.5,
+ 503,
+ 565.0
+ ],
+ "spans": [],
+ "index": 44
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 364,
+ 574,
+ 504,
+ 590
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 364,
+ 574,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 364,
+ 574,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Earthquake: Learned edges at three dif-",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 363,
+ 581,
+ 434,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 363,
+ 581,
+ 434,
+ 591
+ ],
+ "score": 1.0,
+ "content": "ferent stages of training.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 45.5
+ }
+ ],
+ "index": 44.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 655,
+ 356,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 654,
+ 356,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 654,
+ 356,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Results The model can successfully recover the correct edge",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 357,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 357,
+ 678
+ ],
+ "score": 1.0,
+ "content": "structure for all BnLearn graphs considered up to and including",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 357,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 357,
+ 689
+ ],
+ "score": 1.0,
+ "content": "8-variable Asia. Figures 3 and 4 illustrate what the model",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 357,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 357,
+ 700
+ ],
+ "score": 1.0,
+ "content": "has learned at several stages of learning. In these figures, the",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 106,
+ 698,
+ 357,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 133,
+ 711
+ ],
+ "score": 0.93,
+ "content": "\\sigma ( \\gamma _ { i j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 134,
+ 698,
+ 151,
+ 711
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 700,
+ 164,
+ 711
+ ],
+ "score": 0.88,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 698,
+ 357,
+ 711
+ ],
+ "score": 1.0,
+ "content": "adjacency matrix elements are plotted as colored",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 357,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 357,
+ 722
+ ],
+ "score": 1.0,
+ "content": "squares and dots. Off-diagonal terms are unknown, and appear",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 138,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 138,
+ 733
+ ],
+ "score": 1.0,
+ "content": "yellow.",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ }
+ ],
+ "index": 52
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 702
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 702
+ ],
+ "group_id": 2,
+ "lines": [
+ {
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 702
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "1f0d4465e558dd224c900b27c6e12284050ff546dc987c0d1ce52780e5de9750.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 56.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 679.5
+ ],
+ "spans": [],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 366,
+ 679.5,
+ 502,
+ 702.0
+ ],
+ "spans": [],
+ "index": 57
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 364,
+ 711,
+ 504,
+ 727
+ ],
+ "group_id": 2,
+ "lines": [
+ {
+ "bbox": [
+ 363,
+ 709,
+ 505,
+ 720
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 363,
+ 709,
+ 505,
+ 720
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Asia: Learned edges at three different",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 363,
+ 718,
+ 416,
+ 728
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 363,
+ 718,
+ 416,
+ 728
+ ],
+ "score": 1.0,
+ "content": "stages of training.",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ }
+ ],
+ "index": 58.5
+ }
+ ],
+ "index": 57.5
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "7",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 82,
+ 316,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 317,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 317,
+ 95
+ ],
+ "score": 1.0,
+ "content": "5 EXPERIMENTAL SETUP AND RESULTS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 505,
+ 112
+ ],
+ "score": 1.0,
+ "content": "Our experiments aim to evaluate the proposed method to recover the correct causal structure and",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 123
+ ],
+ "score": 1.0,
+ "content": "which elements of the method matter. We first evaluate our model on a synthetic dataset where we",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": "have control over the number of variables and causal edges in the ground-truth SCM. This allows",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 473,
+ 146
+ ],
+ "score": 1.0,
+ "content": "us to verify after learning with what accuracy we recover the individual decisions about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 134,
+ 486,
+ 145
+ ],
+ "score": 0.87,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 132,
+ 506,
+ 146
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "understand the performance of our algorithm under various conditions. We then evaluate our method",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 154,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 154,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": "on real world datasets collected from the BnLearn dataset repository, and show that the proposed",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "approach recovers the true causal structure. We then show that the trained models can correctly",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "predict the consequences of previously unseen interventions on the rest of the graph. We also perform",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 188,
+ 373,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 188,
+ 373,
+ 200
+ ],
+ "score": 1.0,
+ "content": "ablations showing how important is each component of the model.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 105,
+ 100,
+ 506,
+ 200
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 212,
+ 228,
+ 222
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 211,
+ 229,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 211,
+ 229,
+ 224
+ ],
+ "score": 1.0,
+ "content": "5.1 SYNTHETIC DATASETS",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 267
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 267
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 267
+ ],
+ "score": 0.964,
+ "type": "image",
+ "image_path": "f99f187fc376946f89ba5b065c406dc52459b277e1a3d71f66784cd8ed754b6a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 161,
+ 225,
+ 447,
+ 239.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 161,
+ 239.0,
+ 447,
+ 253.0
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 161,
+ 253.0,
+ 447,
+ 267.0
+ ],
+ "spans": [],
+ "index": 13
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 275,
+ 504,
+ 293
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Learned edges at three different stages of training. Left: Chain graph with 4 variables. Right: Fully-connected DAG graph with 4",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 283,
+ 136,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 283,
+ 136,
+ 292
+ ],
+ "score": 1.0,
+ "content": "variables.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14.5
+ }
+ ],
+ "index": 13.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 296,
+ 505,
+ 395
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 295,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "We first evaluate the model’s performance on several randomly-initialized SCMs with specific,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 252,
+ 320
+ ],
+ "score": 1.0,
+ "content": "representative graph structures. For",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 253,
+ 307,
+ 283,
+ 317
+ ],
+ "score": 0.91,
+ "content": "M = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 306,
+ 506,
+ 320
+ ],
+ "score": 1.0,
+ "content": "-variable DAGs, we consider every possible connected",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "graph: chain3, fork3, collider3 and confounder3 (See Fig. 7 in appendix). They exhibit",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "every graph sub-structure that can exist in larger graphs, and must be mastered before tackling larger",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 340,
+ 506,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 506,
+ 353
+ ],
+ "score": 1.0,
+ "content": "graphs. Since the number of possible DAGs grows super-exponentially with the number of variables,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 351,
+ 506,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 120,
+ 363
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 121,
+ 351,
+ 151,
+ 361
+ ],
+ "score": 0.88,
+ "content": "M > 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 351,
+ 506,
+ 363
+ ],
+ "score": 1.0,
+ "content": "up to 8 a selection of representative and edge-case graphs are chosen. The chainM and",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 361,
+ 170,
+ 374
+ ],
+ "score": 1.0,
+ "content": "fullM graphs",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 362,
+ 204,
+ 372
+ ],
+ "score": 0.85,
+ "content": "( \\mathbb { M } = 3 - 8",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 361,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": ") are the minimally and maximally connected M-variable graphs, while the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 506,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 506,
+ 385
+ ],
+ "score": 1.0,
+ "content": "remaining graphs are randomly generated with a varying sparsity level (1-4 expected number of",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 402,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 402,
+ 396
+ ],
+ "score": 1.0,
+ "content": "parents per node). The details of the setup can be found in Appendix A.5.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 105,
+ 295,
+ 506,
+ 396
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 406,
+ 505,
+ 484
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 406,
+ 506,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 406,
+ 506,
+ 419
+ ],
+ "score": 1.0,
+ "content": "Results The model can successfully recover the correct edge structure for all synthetic graphs con-",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "sidered. The learning curves plotting the average cross-entropy (CE) loss for the learned edges against",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 219,
+ 441
+ ],
+ "score": 1.0,
+ "content": "the ground-truth model for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 220,
+ 429,
+ 252,
+ 439
+ ],
+ "score": 0.9,
+ "content": "M = 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 428,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "are shown in Figure 7. The fully-connected confounder3",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 452
+ ],
+ "score": 1.0,
+ "content": "graph is particularly easy to learn, but all 3-variable graphs are learned perfectly. Plots of the same",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "process on 4- through 8-variable graphs were similarly encouraging, with all models converging to",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 506,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 506,
+ 475
+ ],
+ "score": 1.0,
+ "content": "a negligible loss. The results are, however, sensitive to some hyperparameters, notably the DAG",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 237,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 237,
+ 486
+ ],
+ "score": 1.0,
+ "content": "penalty and the sparsity penalty.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 105,
+ 406,
+ 506,
+ 486
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 496,
+ 289,
+ 507
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 291,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 291,
+ 509
+ ],
+ "score": 1.0,
+ "content": "5.2 REAL-WORLD DATASETS: BNLEARN",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 517,
+ 357,
+ 594
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 357,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 357,
+ 529
+ ],
+ "score": 1.0,
+ "content": "The Bayesian Network Repository is a collection of commonly-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 528,
+ 357,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 357,
+ 539
+ ],
+ "score": 1.0,
+ "content": "used causal Bayesian networks from the literature, suitable for",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 539,
+ 357,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 539,
+ 357,
+ 551
+ ],
+ "score": 1.0,
+ "content": "Bayesian and causal learning benchmarks. We evaluate our",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 549,
+ 357,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 357,
+ 562
+ ],
+ "score": 1.0,
+ "content": "model on the Earthquake (Korb & Nicholson, 2010), Cancer",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 560,
+ 357,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 560,
+ 357,
+ 573
+ ],
+ "score": 1.0,
+ "content": "(Korb & Nicholson, 2010) and Asia (Lauritzen & Spiegelhalter,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 571,
+ 358,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 571,
+ 172,
+ 584
+ ],
+ "score": 1.0,
+ "content": "1988) datasets (",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 572,
+ 201,
+ 582
+ ],
+ "score": 0.78,
+ "content": "M = 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 571,
+ 358,
+ 584
+ ],
+ "score": 1.0,
+ "content": ", 5 and 8-variables respectively, max-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 358,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 358,
+ 597
+ ],
+ "score": 1.0,
+ "content": "imum 2 parents per node) in the BnLearn dataset repository.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 517,
+ 358,
+ 597
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 595,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "The ground-truth SCM is given for each dataset, and the functional parameters are represented as",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "conditional probability tables (CPTs). We learn a near-ground-truth MLP from the dataset’s CPT and",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 616,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 616,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "use it as the ground-truth data generator. We also insert a (greater than 1) temperature factor in order",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 104,
+ 625,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 625,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "to increase the likelihood of sampling some very rare events in the CPTs. Details of the setup can be",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 212,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 212,
+ 650
+ ],
+ "score": 1.0,
+ "content": "found in Appendix A.5.1.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 104,
+ 594,
+ 506,
+ 650
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 565
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 565
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 565
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "d9b5c523ea9b4196e4b0c1585be866e7e439433e299ccc6ecc2da33b588016ef.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 365,
+ 520,
+ 503,
+ 542.5
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 365,
+ 542.5,
+ 503,
+ 565.0
+ ],
+ "spans": [],
+ "index": 44
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 364,
+ 574,
+ 504,
+ 590
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 364,
+ 574,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 364,
+ 574,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Earthquake: Learned edges at three dif-",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 363,
+ 581,
+ 434,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 363,
+ 581,
+ 434,
+ 591
+ ],
+ "score": 1.0,
+ "content": "ferent stages of training.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 45.5
+ }
+ ],
+ "index": 44.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 655,
+ 356,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 654,
+ 356,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 654,
+ 356,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Results The model can successfully recover the correct edge",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 357,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 357,
+ 678
+ ],
+ "score": 1.0,
+ "content": "structure for all BnLearn graphs considered up to and including",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 357,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 357,
+ 689
+ ],
+ "score": 1.0,
+ "content": "8-variable Asia. Figures 3 and 4 illustrate what the model",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 357,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 357,
+ 700
+ ],
+ "score": 1.0,
+ "content": "has learned at several stages of learning. In these figures, the",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 106,
+ 698,
+ 357,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 133,
+ 711
+ ],
+ "score": 0.93,
+ "content": "\\sigma ( \\gamma _ { i j } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 134,
+ 698,
+ 151,
+ 711
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 151,
+ 700,
+ 164,
+ 711
+ ],
+ "score": 0.88,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 698,
+ 357,
+ 711
+ ],
+ "score": 1.0,
+ "content": "adjacency matrix elements are plotted as colored",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 357,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 357,
+ 722
+ ],
+ "score": 1.0,
+ "content": "squares and dots. Off-diagonal terms are unknown, and appear",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 138,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 138,
+ 733
+ ],
+ "score": 1.0,
+ "content": "yellow.",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ }
+ ],
+ "index": 52,
+ "bbox_fs": [
+ 105,
+ 654,
+ 357,
+ 733
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 702
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 702
+ ],
+ "group_id": 2,
+ "lines": [
+ {
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 702
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "1f0d4465e558dd224c900b27c6e12284050ff546dc987c0d1ce52780e5de9750.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 56.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 366,
+ 657,
+ 502,
+ 679.5
+ ],
+ "spans": [],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 366,
+ 679.5,
+ 502,
+ 702.0
+ ],
+ "spans": [],
+ "index": 57
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 364,
+ 711,
+ 504,
+ 727
+ ],
+ "group_id": 2,
+ "lines": [
+ {
+ "bbox": [
+ 363,
+ 709,
+ 505,
+ 720
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 363,
+ 709,
+ 505,
+ 720
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Asia: Learned edges at three different",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 363,
+ 718,
+ 416,
+ 728
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 363,
+ 718,
+ 416,
+ 728
+ ],
+ "score": 1.0,
+ "content": "stages of training.",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ }
+ ],
+ "index": 58.5
+ }
+ ],
+ "index": 57.5
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 129,
+ 71,
+ 483,
+ 160
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 129,
+ 71,
+ 483,
+ 160
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 71,
+ 483,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 71,
+ 483,
+ 160
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "4233c6c69310a6d99481788ea1053d2c1bf7319a29ae8964709cdf0d3f64ee2c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 129,
+ 71,
+ 483,
+ 100.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 129,
+ 100.66666666666667,
+ 483,
+ 130.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 129,
+ 130.33333333333334,
+ 483,
+ 160.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 168,
+ 504,
+ 193
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 504,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 487,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Left: Cross entropy (CE) for edge probability between learned and ground-truth graphs for all 3-variable SCMs. Error bars are",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 488,
+ 169,
+ 504,
+ 176
+ ],
+ "score": 0.87,
+ "content": "\\pm 1 \\sigma",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 175,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 175,
+ 506,
+ 186
+ ],
+ "score": 1.0,
+ "content": "over PRNG seeds 1-5. Middle: Edge CE loss for the chain graph with 4-7 variables. Right: Edge CE loss for 3-variable graphs with no",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 185,
+ 298,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 185,
+ 298,
+ 194
+ ],
+ "score": 1.0,
+ "content": "dropout during pretraining, showing the importance of this dropout.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 188,
+ 205,
+ 422,
+ 293
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 188,
+ 205,
+ 422,
+ 293
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 188,
+ 205,
+ 422,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 188,
+ 205,
+ 422,
+ 293
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "62682ff61ea32624cedad428b27f87926c737a7dfde75025a4c006be7dbb5857.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 188,
+ 205,
+ 422,
+ 217.57142857142858
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 188,
+ 217.57142857142858,
+ 422,
+ 230.14285714285717
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 188,
+ 230.14285714285717,
+ 422,
+ 242.71428571428575
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 188,
+ 242.71428571428575,
+ 422,
+ 255.28571428571433
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 188,
+ 255.28571428571433,
+ 422,
+ 267.8571428571429
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 188,
+ 267.8571428571429,
+ 422,
+ 280.42857142857144
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 188,
+ 280.42857142857144,
+ 422,
+ 293.0
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 302,
+ 504,
+ 342
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 301,
+ 505,
+ 311
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Ablations study results on all possible 3 variable graphs. Both graphs show the cross-entropy loss on learned vs ground-truth",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 506,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 506,
+ 319
+ ],
+ "score": 1.0,
+ "content": "edges over training time. Left: Models that infer the intervention (prediction, bold) vs models that have knowledge of the true intervention",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 327
+ ],
+ "score": 1.0,
+ "content": "(ground truth, long dash) vs models that use no knowledge of the intervention at all (no prediction, short dash). Result suggests inferring the",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 335
+ ],
+ "score": 1.0,
+ "content": "intervention works almost as well as knowing the true intervention. Right: Comparisons of model trained with and without DAG regularizer",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 108,
+ 334,
+ 271,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 334,
+ 128,
+ 342
+ ],
+ "score": 0.79,
+ "content": "( L _ { \\mathrm { D A G } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 128,
+ 334,
+ 271,
+ 343
+ ],
+ "score": 1.0,
+ "content": ", showing that DAG regularizer helps convergence.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 359,
+ 505,
+ 414
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 359,
+ 356,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 359,
+ 356,
+ 370
+ ],
+ "score": 1.0,
+ "content": "The color of the squares indicates belief in the presence or",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 370,
+ 505,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 185,
+ 383
+ ],
+ "score": 1.0,
+ "content": "absence of an edge",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 370,
+ 218,
+ 382
+ ],
+ "score": 0.91,
+ "content": "( \\sigma ( \\gamma _ { i j } ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 218,
+ 370,
+ 431,
+ 383
+ ],
+ "score": 1.0,
+ "content": ", while the color of the dot indicates the ground truth",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 431,
+ 371,
+ 444,
+ 382
+ ],
+ "score": 0.86,
+ "content": "c _ { i j }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 444,
+ 370,
+ 505,
+ 383
+ ],
+ "score": 1.0,
+ "content": ". Red indicates",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 380,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 380,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "(belief in) an edge; Blue indicates (belief in) the absence of an edge. Yellow indicates maximum",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 391,
+ 506,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 506,
+ 405
+ ],
+ "score": 1.0,
+ "content": "uncertainty At the beginning of training, the main diagonal is a priori known to be clear of edges,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 403,
+ 215,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 215,
+ 415
+ ],
+ "score": 1.0,
+ "content": "and therefore is solid blue.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 420,
+ 505,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 432
+ ],
+ "score": 1.0,
+ "content": "As training progresses, the beliefs approach the ground truth, which visually appears as the squares",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "converging towards the color of the dot within them. When they coincide, the dot vanishes. An",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 454
+ ],
+ "score": 1.0,
+ "content": "erroneous belief stands out as a red dot on blue square or, vice-versa, a blue dot on red square.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 453,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 453,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "Because we pre-sort the nodes in BnLearn datasets so that they are in topological order, the model",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 477
+ ],
+ "score": 1.0,
+ "content": "must learn a lower triangle. This corresponds to a completely blue uppper triangle. Anti-causal",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 474,
+ 408,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 408,
+ 487
+ ],
+ "score": 1.0,
+ "content": "violations are easily recognizable as non-blue squares in the upper triangle.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 499,
+ 502,
+ 533
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 498,
+ 504,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 498,
+ 504,
+ 512
+ ],
+ "score": 1.0,
+ "content": "Baseline comparisons We compared our method to ICP (Peters et al., 2016) and Eaton & Murphy",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 510,
+ 505,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 505,
+ 522
+ ],
+ "score": 1.0,
+ "content": "(2007a). Eaton & Murphy (2007a) handles uncertain interventions and Peters et al. (2016) handles",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 522,
+ 414,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 522,
+ 414,
+ 534
+ ],
+ "score": 1.0,
+ "content": "unknown interventions. However, neither attempt to predict the intervention.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 117,
+ 564,
+ 495,
+ 599
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 538,
+ 504,
+ 556
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 505,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 505,
+ 549
+ ],
+ "score": 1.0,
+ "content": "Table 1: Baseline comparisons: Cross entropy (lower is better) for edge probability on learned and ground-truth edges on Asia graph.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 546,
+ 351,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 546,
+ 351,
+ 556
+ ],
+ "score": 1.0,
+ "content": "compared to to Peters et al. (2016), (Eaton & Murphy, 2007a) and (Zheng et al., 2018)",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 117,
+ 564,
+ 495,
+ 599
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 117,
+ 564,
+ 495,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 564,
+ 495,
+ 599
+ ],
+ "score": 0.973,
+ "html": "| Algorithm1Training Algorithm |
| 1: procedure TRAINING(Categorical Distribution D,with M nodes and N categories) Let ian integer from O to M-1 |
| 2: 34 | |
| for kpretrain steps do |
| 5: | x~D |
| 6: | c ~ Ber(σ(γ)) |
| 7: 8: | L = -log P(xlc) > Compute log-probability of data given config |
| 0slow ←Adam(0slow,VL) |
| 9: for kintervention steps do | |
| 10: | I_N←randint(O，M-1) |
| 11: | Dint := D with intervention on node I_N |
| 12: | if predicting intervention then |
| 13: | Li←O∀i |
| 14: | for Kpredict steps do |
| 15: | x~Dint |
| 16: | c ~ Ber(σ(γ)) |
| 17: 18: | Li←Li+-logPi(xlci;Oslow)∀i |
| >Accumulate NLL for every node i separately I_N←argmax(Li) |
| |
| 19: 20: | gammagrads，logregrets =[]，[] >Transfer Episode Adaptation Loop |
| 21: | for Kepisode steps do x~Dint |
| 22: | gammagrad，logregret = 0，0 |
| 23: | for kcfg steps do |
| 24: | c ~ Ber(σ(γ)) |
| 25: | Li=-log Pi(xlci;0slow）∀i |
| 26: | gammagrad += σ(γ)-c |
| 27: | logregret += ∑ Li |
| 28: | I_N |
| 29: | gammagrads.append(gammagrad) logregrets.append(logregret) |
| |
| 30: | J ←入MaxEnt LMaxEnt(γ)+ 入sparse LSparse(γ)+ 入DAG LDAG(γ) |
| 31: | Vγ←VγJ+∑gammagradskijlogregrets.softmax(0) ki |
| 32: | k γ ← Adam(γ,∀γ) |
",
+ "type": "table",
+ "image_path": "fd733ccb41b58584ef493bdf1ebe59b58d01aca69982b90e1a69c2ae8f2517a9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 124,
+ 506,
+ 257.33333333333337
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 257.33333333333337,
+ 506,
+ 390.66666666666674
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 390.66666666666674,
+ 506,
+ 524.0000000000001
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 545,
+ 202,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 203,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 203,
+ 558
+ ],
+ "score": 1.0,
+ "content": "A.3 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 560,
+ 505,
+ 692
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "Interventions In a purely-observational setting, it is known that causal graphs can be distinguished",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 571,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 571,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "only up to a Markov equivalence class. In order to identify the true causal graph intervention data is",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 594
+ ],
+ "score": 1.0,
+ "content": "needed (Eberhardt et al., 2012). Several types of common interventions may be available (Eaton &",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 593,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 593,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "Murphy, 2007a). These are: No intervention: only observational data is obtained from the ground",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 604,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 604,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "truth causal model. Hard/perfect: the value of a single or several variables is fixed and then ancestral",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 627
+ ],
+ "score": 1.0,
+ "content": "sampling is performed on the other variables. Soft/imperfect: the conditional distribution of the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 638
+ ],
+ "score": 1.0,
+ "content": "variable on which the intervention is performed is changed. Uncertain: the learner is not sure of",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": "which variable exactly the intervention affected directly. Here we make use of soft interventions for",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 660
+ ],
+ "score": 1.0,
+ "content": "several reasons: First, they include hard interventions as a limiting case and hence are more general.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 659,
+ 506,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 659,
+ 506,
+ 671
+ ],
+ "score": 1.0,
+ "content": "Second, in many real-world scenarios, it is more difficult to perform a hard intervention compared to",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 670,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "a soft one. We also deal with a special case of uncertain interventions, where the variable selected for",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 470,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 470,
+ 693
+ ],
+ "score": 1.0,
+ "content": "intervention is random and unknown. We call these unidentified or unknown interventions.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Causal sufficiency The inability to distinguish which causal graph, within a Markov equivalence",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "class, is the correct one in the purely-observational setting is called the identifiability problem. In",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 506,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 506,
+ 733
+ ],
+ "score": 1.0,
+ "content": "our setting, all variables are observed (there are no latent confounders) and all interventions are",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "page_idx": 13,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 234,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 235,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 235,
+ 94
+ ],
+ "score": 1.0,
+ "content": "A.2 TRAINING ALGORITHM",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 103,
+ 343,
+ 114
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 102,
+ 343,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 102,
+ 343,
+ 115
+ ],
+ "score": 1.0,
+ "content": "In this section, we describe the training algorithm in detail.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 106,
+ 102,
+ 343,
+ 115
+ ]
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 105,
+ 124,
+ 506,
+ 524
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 105,
+ 124,
+ 506,
+ 524
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 124,
+ 506,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 124,
+ 506,
+ 524
+ ],
+ "score": 0.98,
+ "html": "| Algorithm1Training Algorithm |
| 1: procedure TRAINING(Categorical Distribution D,with M nodes and N categories) Let ian integer from O to M-1 |
| 2: 34 | |
| for kpretrain steps do |
| 5: | x~D |
| 6: | c ~ Ber(σ(γ)) |
| 7: 8: | L = -log P(xlc) > Compute log-probability of data given config |
| 0slow ←Adam(0slow,VL) |
| 9: for kintervention steps do | |
| 10: | I_N←randint(O，M-1) |
| 11: | Dint := D with intervention on node I_N |
| 12: | if predicting intervention then |
| 13: | Li←O∀i |
| 14: | for Kpredict steps do |
| 15: | x~Dint |
| 16: | c ~ Ber(σ(γ)) |
| 17: 18: | Li←Li+-logPi(xlci;Oslow)∀i |
| >Accumulate NLL for every node i separately I_N←argmax(Li) |
| |
| 19: 20: | gammagrads，logregrets =[]，[] >Transfer Episode Adaptation Loop |
| 21: | for Kepisode steps do x~Dint |
| 22: | gammagrad，logregret = 0，0 |
| 23: | for kcfg steps do |
| 24: | c ~ Ber(σ(γ)) |
| 25: | Li=-log Pi(xlci;0slow）∀i |
| 26: | gammagrad += σ(γ)-c |
| 27: | logregret += ∑ Li |
| 28: | I_N |
| 29: | gammagrads.append(gammagrad) logregrets.append(logregret) |
| |
| 30: | J ←入MaxEnt LMaxEnt(γ)+ 入sparse LSparse(γ)+ 入DAG LDAG(γ) |
| 31: | Vγ←VγJ+∑gammagradskijlogregrets.softmax(0) ki |
| 32: | k γ ← Adam(γ,∀γ) |
",
+ "type": "table",
+ "image_path": "fd733ccb41b58584ef493bdf1ebe59b58d01aca69982b90e1a69c2ae8f2517a9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 124,
+ 506,
+ 257.33333333333337
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 257.33333333333337,
+ 506,
+ 390.66666666666674
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 390.66666666666674,
+ 506,
+ 524.0000000000001
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 545,
+ 202,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 203,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 203,
+ 558
+ ],
+ "score": 1.0,
+ "content": "A.3 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 560,
+ 505,
+ 692
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "Interventions In a purely-observational setting, it is known that causal graphs can be distinguished",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 571,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 571,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "only up to a Markov equivalence class. In order to identify the true causal graph intervention data is",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 594
+ ],
+ "score": 1.0,
+ "content": "needed (Eberhardt et al., 2012). Several types of common interventions may be available (Eaton &",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 593,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 593,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "Murphy, 2007a). These are: No intervention: only observational data is obtained from the ground",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 604,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 604,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "truth causal model. Hard/perfect: the value of a single or several variables is fixed and then ancestral",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 627
+ ],
+ "score": 1.0,
+ "content": "sampling is performed on the other variables. Soft/imperfect: the conditional distribution of the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 638
+ ],
+ "score": 1.0,
+ "content": "variable on which the intervention is performed is changed. Uncertain: the learner is not sure of",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": "which variable exactly the intervention affected directly. Here we make use of soft interventions for",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 660
+ ],
+ "score": 1.0,
+ "content": "several reasons: First, they include hard interventions as a limiting case and hence are more general.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 659,
+ 506,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 659,
+ 506,
+ 671
+ ],
+ "score": 1.0,
+ "content": "Second, in many real-world scenarios, it is more difficult to perform a hard intervention compared to",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 670,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "a soft one. We also deal with a special case of uncertain interventions, where the variable selected for",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 470,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 470,
+ 693
+ ],
+ "score": 1.0,
+ "content": "intervention is random and unknown. We call these unidentified or unknown interventions.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 560,
+ 506,
+ 693
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Causal sufficiency The inability to distinguish which causal graph, within a Markov equivalence",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "class, is the correct one in the purely-observational setting is called the identifiability problem. In",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 506,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 506,
+ 733
+ ],
+ "score": 1.0,
+ "content": "our setting, all variables are observed (there are no latent confounders) and all interventions are",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "random and independent. Hence, within our setting the true causal graph is always identifiable in",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": "principle (Eberhardt et al., 2012; Heinze-Deml et al., 2018). We consider here situations where a",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "single variable is randomly selected and intervened upon with a soft or imprecise intervention, its",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 275,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 275,
+ 127
+ ],
+ "score": 1.0,
+ "content": "identity is unknown and must be inferred.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 698,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 126
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "random and independent. Hence, within our setting the true causal graph is always identifiable in",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": "principle (Eberhardt et al., 2012; Heinze-Deml et al., 2018). We consider here situations where a",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "single variable is randomly selected and intervened upon with a soft or imprecise intervention, its",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 275,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 275,
+ 127
+ ],
+ "score": 1.0,
+ "content": "identity is unknown and must be inferred.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 1.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 144,
+ 231,
+ 155
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 231,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 231,
+ 155
+ ],
+ "score": 1.0,
+ "content": "A.4 EXPERIMENTAL SETUP",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 165,
+ 504,
+ 188
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "For all datasets, the weight parameters for the learned model is initialized randomly. In order to not",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 175,
+ 440,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 175,
+ 240,
+ 190
+ ],
+ "score": 1.0,
+ "content": "bias the structural parameters, all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 178,
+ 248,
+ 188
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 175,
+ 440,
+ 190
+ ],
+ "score": 1.0,
+ "content": "is initialized to 0.5 in the beginning of training.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 204,
+ 208,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 209,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 209,
+ 216
+ ],
+ "score": 1.0,
+ "content": "A.5 SYNTHETIC DATA",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 226,
+ 505,
+ 270
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 150,
+ 238
+ ],
+ "score": 1.0,
+ "content": "SCM with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 228,
+ 157,
+ 236
+ ],
+ "score": 0.72,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 226,
+ 254,
+ 238
+ ],
+ "score": 1.0,
+ "content": "variables is modeled by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 228,
+ 262,
+ 236
+ ],
+ "score": 0.73,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 226,
+ 505,
+ 238
+ ],
+ "score": 1.0,
+ "content": "feedforward neural networks (MLPs) as described in section",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 250
+ ],
+ "score": 1.0,
+ "content": "3.1. For simplicity, we assume use an acyclic causal graph such that we could easily sample from",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 248,
+ 505,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 297,
+ 261
+ ],
+ "score": 1.0,
+ "content": "it. Hence, given any pair of random variables",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 298,
+ 249,
+ 306,
+ 258
+ ],
+ "score": 0.75,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 248,
+ 325,
+ 261
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 249,
+ 334,
+ 258
+ ],
+ "score": 0.81,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 248,
+ 365,
+ 261
+ ],
+ "score": 1.0,
+ "content": ", either",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 249,
+ 399,
+ 259
+ ],
+ "score": 0.88,
+ "content": "A B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 248,
+ 403,
+ 261
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 249,
+ 438,
+ 259
+ ],
+ "score": 0.88,
+ "content": "B A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 248,
+ 451,
+ 261
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 451,
+ 249,
+ 460,
+ 258
+ ],
+ "score": 0.76,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 460,
+ 248,
+ 479,
+ 261
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 479,
+ 249,
+ 488,
+ 258
+ ],
+ "score": 0.81,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 489,
+ 248,
+ 505,
+ 261
+ ],
+ "score": 1.0,
+ "content": "are",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 160,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 160,
+ 272
+ ],
+ "score": 1.0,
+ "content": "independent.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 276,
+ 505,
+ 309
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 344,
+ 289
+ ],
+ "score": 1.0,
+ "content": "The MLP representing the ground-truth SCM has its weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 277,
+ 351,
+ 286
+ ],
+ "score": 0.81,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 352,
+ 275,
+ 505,
+ 289
+ ],
+ "score": 1.0,
+ "content": "initialized use orthogonal initialization",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 287,
+ 505,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 422,
+ 299
+ ],
+ "score": 1.0,
+ "content": "with gain 2.5 and the biases are initialized using a uniform initialization between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 288,
+ 444,
+ 298
+ ],
+ "score": 0.8,
+ "content": "- 1 . 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 444,
+ 287,
+ 505,
+ 299
+ ],
+ "score": 1.0,
+ "content": "and 1.1, which",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 400,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 400,
+ 311
+ ],
+ "score": 1.0,
+ "content": "was empirically found to yield \"interesting\" yet learnable random SCMs.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 428
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 428
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 428
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "3a790701f5c12dec0fe6dde4d2cd704750cc5a7f50499f417a7dd1e4510253c9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 365.3333333333333
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 113,
+ 365.3333333333333,
+ 460,
+ 396.66666666666663
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 113,
+ 396.66666666666663,
+ 460,
+ 427.99999999999994
+ ],
+ "spans": [],
+ "index": 17
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 440,
+ 504,
+ 457
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 450
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Left: Every possible 3-variable connected DAG. Right: Cross entropy for edge probability between learned and ground-truth SCM",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 176,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 176,
+ 457
+ ],
+ "score": 1.0,
+ "content": "for all 3-variable SCMs.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5
+ }
+ ],
+ "index": 17.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 482,
+ 265,
+ 493
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 481,
+ 267,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 481,
+ 267,
+ 494
+ ],
+ "score": 1.0,
+ "content": "A.5.1 BNLEARN DATA REPOSITORY",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 501,
+ 506,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "The repo contains many datasets with various sizes and structures modeling different variables. We",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 512,
+ 507,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 512,
+ 507,
+ 527
+ ],
+ "score": 1.0,
+ "content": "evaluate our model on 3 of the datasets in the repo, namely the Earthquake (Korb & Nicholson,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "2010), Cancer (Korb & Nicholson, 2010) and Asia (Lauritzen & Spiegelhalter, 1988) datasets. The",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 534,
+ 506,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 534,
+ 506,
+ 547
+ ],
+ "score": 1.0,
+ "content": "ground-truth model structure for the Cancer (Korb & Nicholson, 2010) and Earthquake (Korb &",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 545,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "Nicholson, 2010) datasets are shown in Figure 8. Note that even though the structure for the 2 datasets",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 557,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "seems to be the same, the conditional probability tables (CPTs) for these 2 datasets are very different",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 567,
+ 433,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 567,
+ 433,
+ 580
+ ],
+ "score": 1.0,
+ "content": "and hence results in different structured causal models (SCMs) for the 2 datasets.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 711
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 711
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 711
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "74b53bb5cabd72d9b02acd007967754ee13b35decbfe1cd63be56773fcce91d8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 29,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 637.6666666666666
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 160,
+ 637.6666666666666,
+ 451,
+ 674.3333333333333
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 160,
+ 674.3333333333333,
+ 451,
+ 710.9999999999999
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 128,
+ 720,
+ 482,
+ 730
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 720,
+ 482,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 720,
+ 482,
+ 730
+ ],
+ "score": 1.0,
+ "content": "Figure 8: Left: Ground Truth SCM for Cancer. Middle: Groundtruth SCM for Earthquake. Right: Groundtruth SCM for Asia.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.0
+ }
+ ],
+ "page_idx": 14,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 126
+ ],
+ "lines": [],
+ "index": 1.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 127
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 144,
+ 231,
+ 155
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 231,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 144,
+ 231,
+ 155
+ ],
+ "score": 1.0,
+ "content": "A.4 EXPERIMENTAL SETUP",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 165,
+ 504,
+ 188
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "For all datasets, the weight parameters for the learned model is initialized randomly. In order to not",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 175,
+ 440,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 175,
+ 240,
+ 190
+ ],
+ "score": 1.0,
+ "content": "bias the structural parameters, all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 178,
+ 248,
+ 188
+ ],
+ "score": 0.82,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 175,
+ 440,
+ 190
+ ],
+ "score": 1.0,
+ "content": "is initialized to 0.5 in the beginning of training.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 164,
+ 505,
+ 190
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 204,
+ 208,
+ 216
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 209,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 209,
+ 216
+ ],
+ "score": 1.0,
+ "content": "A.5 SYNTHETIC DATA",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 226,
+ 505,
+ 270
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 150,
+ 238
+ ],
+ "score": 1.0,
+ "content": "SCM with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 228,
+ 157,
+ 236
+ ],
+ "score": 0.72,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 226,
+ 254,
+ 238
+ ],
+ "score": 1.0,
+ "content": "variables is modeled by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 228,
+ 262,
+ 236
+ ],
+ "score": 0.73,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 226,
+ 505,
+ 238
+ ],
+ "score": 1.0,
+ "content": "feedforward neural networks (MLPs) as described in section",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 250
+ ],
+ "score": 1.0,
+ "content": "3.1. For simplicity, we assume use an acyclic causal graph such that we could easily sample from",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 248,
+ 505,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 248,
+ 297,
+ 261
+ ],
+ "score": 1.0,
+ "content": "it. Hence, given any pair of random variables",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 298,
+ 249,
+ 306,
+ 258
+ ],
+ "score": 0.75,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 248,
+ 325,
+ 261
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 249,
+ 334,
+ 258
+ ],
+ "score": 0.81,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 248,
+ 365,
+ 261
+ ],
+ "score": 1.0,
+ "content": ", either",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 249,
+ 399,
+ 259
+ ],
+ "score": 0.88,
+ "content": "A B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 248,
+ 403,
+ 261
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 249,
+ 438,
+ 259
+ ],
+ "score": 0.88,
+ "content": "B A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 248,
+ 451,
+ 261
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 451,
+ 249,
+ 460,
+ 258
+ ],
+ "score": 0.76,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 460,
+ 248,
+ 479,
+ 261
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 479,
+ 249,
+ 488,
+ 258
+ ],
+ "score": 0.81,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 489,
+ 248,
+ 505,
+ 261
+ ],
+ "score": 1.0,
+ "content": "are",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 259,
+ 160,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 160,
+ 272
+ ],
+ "score": 1.0,
+ "content": "independent.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 226,
+ 506,
+ 272
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 276,
+ 505,
+ 309
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 344,
+ 289
+ ],
+ "score": 1.0,
+ "content": "The MLP representing the ground-truth SCM has its weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 277,
+ 351,
+ 286
+ ],
+ "score": 0.81,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 352,
+ 275,
+ 505,
+ 289
+ ],
+ "score": 1.0,
+ "content": "initialized use orthogonal initialization",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 287,
+ 505,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 422,
+ 299
+ ],
+ "score": 1.0,
+ "content": "with gain 2.5 and the biases are initialized using a uniform initialization between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 288,
+ 444,
+ 298
+ ],
+ "score": 0.8,
+ "content": "- 1 . 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 444,
+ 287,
+ 505,
+ 299
+ ],
+ "score": 1.0,
+ "content": "and 1.1, which",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 400,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 400,
+ 311
+ ],
+ "score": 1.0,
+ "content": "was empirically found to yield \"interesting\" yet learnable random SCMs.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 105,
+ 275,
+ 505,
+ 311
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 428
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 428
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 428
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "3a790701f5c12dec0fe6dde4d2cd704750cc5a7f50499f417a7dd1e4510253c9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 113,
+ 334,
+ 460,
+ 365.3333333333333
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 113,
+ 365.3333333333333,
+ 460,
+ 396.66666666666663
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 113,
+ 396.66666666666663,
+ 460,
+ 427.99999999999994
+ ],
+ "spans": [],
+ "index": 17
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 440,
+ 504,
+ 457
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 439,
+ 506,
+ 450
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Left: Every possible 3-variable connected DAG. Right: Cross entropy for edge probability between learned and ground-truth SCM",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 448,
+ 176,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 448,
+ 176,
+ 457
+ ],
+ "score": 1.0,
+ "content": "for all 3-variable SCMs.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5
+ }
+ ],
+ "index": 17.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 482,
+ 265,
+ 493
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 481,
+ 267,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 481,
+ 267,
+ 494
+ ],
+ "score": 1.0,
+ "content": "A.5.1 BNLEARN DATA REPOSITORY",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 501,
+ 506,
+ 579
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "The repo contains many datasets with various sizes and structures modeling different variables. We",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 512,
+ 507,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 512,
+ 507,
+ 527
+ ],
+ "score": 1.0,
+ "content": "evaluate our model on 3 of the datasets in the repo, namely the Earthquake (Korb & Nicholson,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "2010), Cancer (Korb & Nicholson, 2010) and Asia (Lauritzen & Spiegelhalter, 1988) datasets. The",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 534,
+ 506,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 534,
+ 506,
+ 547
+ ],
+ "score": 1.0,
+ "content": "ground-truth model structure for the Cancer (Korb & Nicholson, 2010) and Earthquake (Korb &",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 545,
+ 506,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 506,
+ 559
+ ],
+ "score": 1.0,
+ "content": "Nicholson, 2010) datasets are shown in Figure 8. Note that even though the structure for the 2 datasets",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 557,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 557,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "seems to be the same, the conditional probability tables (CPTs) for these 2 datasets are very different",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 567,
+ 433,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 567,
+ 433,
+ 580
+ ],
+ "score": 1.0,
+ "content": "and hence results in different structured causal models (SCMs) for the 2 datasets.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 104,
+ 502,
+ 507,
+ 580
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 711
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 711
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 711
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "74b53bb5cabd72d9b02acd007967754ee13b35decbfe1cd63be56773fcce91d8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 29,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 160,
+ 601,
+ 451,
+ 637.6666666666666
+ ],
+ "spans": [],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 160,
+ 637.6666666666666,
+ 451,
+ 674.3333333333333
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 160,
+ 674.3333333333333,
+ 451,
+ 710.9999999999999
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 128,
+ 720,
+ 482,
+ 730
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 720,
+ 482,
+ 730
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 720,
+ 482,
+ 730
+ ],
+ "score": 1.0,
+ "content": "Figure 8: Left: Ground Truth SCM for Cancer. Middle: Groundtruth SCM for Earthquake. Right: Groundtruth SCM for Asia.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.0
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 334,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 336,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 336,
+ 95
+ ],
+ "score": 1.0,
+ "content": "A.5.2 TRAINING GROUND-TRUTH DATA GENERATOR",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 101,
+ 505,
+ 146
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 101,
+ 505,
+ 114
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 101,
+ 505,
+ 114
+ ],
+ "score": 1.0,
+ "content": "Because a CPT is capable of representing any distribution, and MLPs are strictly less powerful in",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 112,
+ 507,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 112,
+ 507,
+ 125
+ ],
+ "score": 1.0,
+ "content": "this respect, it may not be possible to learn perfectly the distribution with our MLP learner model.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 123,
+ 505,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 123,
+ 505,
+ 137
+ ],
+ "score": 1.0,
+ "content": "We therefore train a near-ground-truth MLP to replicate as closely as possible the CPT’s probability",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 134,
+ 420,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 134,
+ 420,
+ 147
+ ],
+ "score": 1.0,
+ "content": "table, and then use this trained MLP are the ground-truth SCM data generator.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 151,
+ 505,
+ 196
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 152,
+ 505,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 152,
+ 505,
+ 164
+ ],
+ "score": 1.0,
+ "content": "Training is by 1000 iterations of full-batch gradient descent with learning rate 0.001 and momentum",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 162,
+ 505,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 162,
+ 408,
+ 175
+ ],
+ "score": 1.0,
+ "content": "0.9, with all possible parent values masked with the ground-truth vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 164,
+ 418,
+ 174
+ ],
+ "score": 0.86,
+ "content": "\\gamma _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 162,
+ 505,
+ 175
+ ],
+ "score": 1.0,
+ "content": ". The objective is to",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 174,
+ 505,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 174,
+ 505,
+ 186
+ ],
+ "score": 1.0,
+ "content": "minimize mean squared error between the MLP’s logits and the log-probability as drawn from the",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 184,
+ 391,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 363,
+ 196
+ ],
+ "score": 1.0,
+ "content": "CPT. If the CPT contains a zero, it is approximated by a logit of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 364,
+ 185,
+ 388,
+ 195
+ ],
+ "score": 0.68,
+ "content": "- 1 0 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 388,
+ 184,
+ 391,
+ 196
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 201,
+ 505,
+ 267
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 506,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 506,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Given that some of the CPTs contain very unlikely events, we have found it necessary to add a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 213,
+ 505,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 213,
+ 505,
+ 225
+ ],
+ "score": 1.0,
+ "content": "temperature parameter in order to make them more frequent. The near-ground-truth MLP model’s",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 224,
+ 504,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 224,
+ 504,
+ 235
+ ],
+ "score": 1.0,
+ "content": "logit outputs are divided by the temperature before being used for sampling. Temperatures above",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "1 result in more uniform distributions for all causal variables; Temperatures below 1 result in less",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 244,
+ 506,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 244,
+ 506,
+ 258
+ ],
+ "score": 1.0,
+ "content": "uniform, sharper distributions that peak around the most likely value. We find empirically that a",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 256,
+ 365,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 365,
+ 268
+ ],
+ "score": 1.0,
+ "content": "temperature of about 2 is required for our BnLearn benchmarks.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 280,
+ 224,
+ 291
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 225,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 225,
+ 293
+ ],
+ "score": 1.0,
+ "content": "A.6 EFFECT OF SPARSITY",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 302,
+ 505,
+ 345
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 143,
+ 315
+ ],
+ "score": 1.0,
+ "content": "We use a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 302,
+ 156,
+ 311
+ ],
+ "score": 0.71,
+ "content": "L 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 300,
+ 308,
+ 315
+ ],
+ "score": 1.0,
+ "content": "regularizer on the structure parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 303,
+ 316,
+ 312
+ ],
+ "score": 0.84,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 300,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "to encourage a sparse representation of edges in",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 312,
+ 505,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 312,
+ 353,
+ 324
+ ],
+ "score": 1.0,
+ "content": "the causal graph. In order to better understand the effect of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 312,
+ 367,
+ 322
+ ],
+ "score": 0.68,
+ "content": "L 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 312,
+ 505,
+ 324
+ ],
+ "score": 1.0,
+ "content": "regularizer, we conducted ablation",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 322,
+ 505,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 164,
+ 336
+ ],
+ "score": 1.0,
+ "content": "studies on the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 323,
+ 178,
+ 333
+ ],
+ "score": 0.67,
+ "content": "L 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 178,
+ 322,
+ 505,
+ 336
+ ],
+ "score": 1.0,
+ "content": "regularizer. It seems that the regularizer has an small effect on rate of converges",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 334,
+ 437,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 437,
+ 346
+ ],
+ "score": 1.0,
+ "content": "and that the model converges faster with the regularizer. This is shown in Figure 9",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 446
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 446
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 446
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "db53278365d2e1de3d89b57ce0a66803a6f2763f2c38f0b3d98a6e5b040f87b2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 390.6666666666667
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 124,
+ 390.6666666666667,
+ 480,
+ 418.33333333333337
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 124,
+ 418.33333333333337,
+ 480,
+ 446.00000000000006
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 198,
+ 453,
+ 413,
+ 462
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 197,
+ 453,
+ 414,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 197,
+ 453,
+ 414,
+ 463
+ ],
+ "score": 1.0,
+ "content": "Figure 9: Effect of Sparsity: On 5 variable, 6 variable and 8 variable Nodes",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 480,
+ 246,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 248,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 248,
+ 492
+ ],
+ "score": 1.0,
+ "content": "A.7 EFFECT OF TEMPERATURE",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "As noted in section A.5.1, we have introduced a temperature hyperparameter in order to encourage",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 511,
+ 506,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 506,
+ 524
+ ],
+ "score": 1.0,
+ "content": "the groundtruth model to generate some very rare events in the conditional probability tables (CPTs)",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "more frequently. We run ablation studies to understand the importance of the temperature term. A",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "temperature of 1 corresponds to no changes to the underlying CPTs. As shown in Figure 12, for the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "score": 1.0,
+ "content": "Cancer (Korb & Nicholson, 2010) dataset, a temperature of 2 improves the accuracy of causal graph",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 146,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 146,
+ 569
+ ],
+ "score": 1.0,
+ "content": "recovery.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 702
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 702
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 702
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "d8967b64efd9a61747efc4636ecfe6584f61e715cdd46ad48ac565ed8ddfd481.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 31.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 645.5
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 225,
+ 645.5,
+ 376,
+ 702.0
+ ],
+ "spans": [],
+ "index": 32
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 133,
+ 710,
+ 475,
+ 719
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 709,
+ 477,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 709,
+ 477,
+ 721
+ ],
+ "score": 1.0,
+ "content": "Figure 10: Cross entropy for edge probability between learned and ground-truth SCM for Cancer at varying temperatures.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.25
+ }
+ ],
+ "page_idx": 15,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 752,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 334,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 336,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 336,
+ 95
+ ],
+ "score": 1.0,
+ "content": "A.5.2 TRAINING GROUND-TRUTH DATA GENERATOR",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 101,
+ 505,
+ 146
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 101,
+ 505,
+ 114
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 101,
+ 505,
+ 114
+ ],
+ "score": 1.0,
+ "content": "Because a CPT is capable of representing any distribution, and MLPs are strictly less powerful in",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 112,
+ 507,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 112,
+ 507,
+ 125
+ ],
+ "score": 1.0,
+ "content": "this respect, it may not be possible to learn perfectly the distribution with our MLP learner model.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 123,
+ 505,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 123,
+ 505,
+ 137
+ ],
+ "score": 1.0,
+ "content": "We therefore train a near-ground-truth MLP to replicate as closely as possible the CPT’s probability",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 134,
+ 420,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 134,
+ 420,
+ 147
+ ],
+ "score": 1.0,
+ "content": "table, and then use this trained MLP are the ground-truth SCM data generator.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 101,
+ 507,
+ 147
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 151,
+ 505,
+ 196
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 152,
+ 505,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 152,
+ 505,
+ 164
+ ],
+ "score": 1.0,
+ "content": "Training is by 1000 iterations of full-batch gradient descent with learning rate 0.001 and momentum",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 162,
+ 505,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 162,
+ 408,
+ 175
+ ],
+ "score": 1.0,
+ "content": "0.9, with all possible parent values masked with the ground-truth vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 164,
+ 418,
+ 174
+ ],
+ "score": 0.86,
+ "content": "\\gamma _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 162,
+ 505,
+ 175
+ ],
+ "score": 1.0,
+ "content": ". The objective is to",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 174,
+ 505,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 174,
+ 505,
+ 186
+ ],
+ "score": 1.0,
+ "content": "minimize mean squared error between the MLP’s logits and the log-probability as drawn from the",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 184,
+ 391,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 184,
+ 363,
+ 196
+ ],
+ "score": 1.0,
+ "content": "CPT. If the CPT contains a zero, it is approximated by a logit of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 364,
+ 185,
+ 388,
+ 195
+ ],
+ "score": 0.68,
+ "content": "- 1 0 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 388,
+ 184,
+ 391,
+ 196
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 152,
+ 505,
+ 196
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 201,
+ 505,
+ 267
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 506,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 506,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Given that some of the CPTs contain very unlikely events, we have found it necessary to add a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 213,
+ 505,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 213,
+ 505,
+ 225
+ ],
+ "score": 1.0,
+ "content": "temperature parameter in order to make them more frequent. The near-ground-truth MLP model’s",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 224,
+ 504,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 224,
+ 504,
+ 235
+ ],
+ "score": 1.0,
+ "content": "logit outputs are divided by the temperature before being used for sampling. Temperatures above",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "1 result in more uniform distributions for all causal variables; Temperatures below 1 result in less",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 244,
+ 506,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 244,
+ 506,
+ 258
+ ],
+ "score": 1.0,
+ "content": "uniform, sharper distributions that peak around the most likely value. We find empirically that a",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 256,
+ 365,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 365,
+ 268
+ ],
+ "score": 1.0,
+ "content": "temperature of about 2 is required for our BnLearn benchmarks.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 201,
+ 506,
+ 268
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 280,
+ 224,
+ 291
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 225,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 225,
+ 293
+ ],
+ "score": 1.0,
+ "content": "A.6 EFFECT OF SPARSITY",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 302,
+ 505,
+ 345
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 143,
+ 315
+ ],
+ "score": 1.0,
+ "content": "We use a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 302,
+ 156,
+ 311
+ ],
+ "score": 0.71,
+ "content": "L 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 300,
+ 308,
+ 315
+ ],
+ "score": 1.0,
+ "content": "regularizer on the structure parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 303,
+ 316,
+ 312
+ ],
+ "score": 0.84,
+ "content": "\\gamma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 300,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "to encourage a sparse representation of edges in",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 312,
+ 505,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 312,
+ 353,
+ 324
+ ],
+ "score": 1.0,
+ "content": "the causal graph. In order to better understand the effect of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 312,
+ 367,
+ 322
+ ],
+ "score": 0.68,
+ "content": "L 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 312,
+ 505,
+ 324
+ ],
+ "score": 1.0,
+ "content": "regularizer, we conducted ablation",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 322,
+ 505,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 164,
+ 336
+ ],
+ "score": 1.0,
+ "content": "studies on the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 323,
+ 178,
+ 333
+ ],
+ "score": 0.67,
+ "content": "L 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 178,
+ 322,
+ 505,
+ 336
+ ],
+ "score": 1.0,
+ "content": "regularizer. It seems that the regularizer has an small effect on rate of converges",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 334,
+ 437,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 437,
+ 346
+ ],
+ "score": 1.0,
+ "content": "and that the model converges faster with the regularizer. This is shown in Figure 9",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 300,
+ 505,
+ 346
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 446
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 446
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 446
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "db53278365d2e1de3d89b57ce0a66803a6f2763f2c38f0b3d98a6e5b040f87b2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 124,
+ 363,
+ 480,
+ 390.6666666666667
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 124,
+ 390.6666666666667,
+ 480,
+ 418.33333333333337
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 124,
+ 418.33333333333337,
+ 480,
+ 446.00000000000006
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 198,
+ 453,
+ 413,
+ 462
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 197,
+ 453,
+ 414,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 197,
+ 453,
+ 414,
+ 463
+ ],
+ "score": 1.0,
+ "content": "Figure 9: Effect of Sparsity: On 5 variable, 6 variable and 8 variable Nodes",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 480,
+ 246,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 248,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 248,
+ 492
+ ],
+ "score": 1.0,
+ "content": "A.7 EFFECT OF TEMPERATURE",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 506,
+ 514
+ ],
+ "score": 1.0,
+ "content": "As noted in section A.5.1, we have introduced a temperature hyperparameter in order to encourage",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 511,
+ 506,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 506,
+ 524
+ ],
+ "score": 1.0,
+ "content": "the groundtruth model to generate some very rare events in the conditional probability tables (CPTs)",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "more frequently. We run ablation studies to understand the importance of the temperature term. A",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "temperature of 1 corresponds to no changes to the underlying CPTs. As shown in Figure 12, for the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 558
+ ],
+ "score": 1.0,
+ "content": "Cancer (Korb & Nicholson, 2010) dataset, a temperature of 2 improves the accuracy of causal graph",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 146,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 146,
+ 569
+ ],
+ "score": 1.0,
+ "content": "recovery.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 500,
+ 506,
+ 569
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 702
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 702
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 702
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "d8967b64efd9a61747efc4636ecfe6584f61e715cdd46ad48ac565ed8ddfd481.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 31.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 225,
+ 589,
+ 376,
+ 645.5
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 225,
+ 645.5,
+ 376,
+ 702.0
+ ],
+ "spans": [],
+ "index": 32
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 133,
+ 710,
+ 475,
+ 719
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 709,
+ 477,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 709,
+ 477,
+ 721
+ ],
+ "score": 1.0,
+ "content": "Figure 10: Cross entropy for edge probability between learned and ground-truth SCM for Cancer at varying temperatures.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.25
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 288
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 288
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 288
+ ],
+ "score": 0.961,
+ "type": "image",
+ "image_path": "93a87e2e595b8857314a6cccac873b3eb129f202923fb425b9075b306cf5a5f5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 208.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 139,
+ 208.66666666666666,
+ 467,
+ 248.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 139,
+ 248.33333333333331,
+ 467,
+ 288.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 105,
+ 298,
+ 506,
+ 315
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "Figure 11: Cross entropy for edge probability between learned and ground-truth SCM. Left: The Earthquake dataset with 6 variables. Right:",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 203,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 203,
+ 316
+ ],
+ "score": 1.0,
+ "content": "The Asia dataset with 8 variables",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 624
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 624
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 624
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "97ce72bf68cf2b4152045f11fbfda91a63a4d62a35f39a1aa1fc0b78630d6cf9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 528.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 150,
+ 528.0,
+ 469,
+ 576.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 150,
+ 576.0,
+ 469,
+ 624.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 103,
+ 633,
+ 505,
+ 651
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "Figure 12: Left: SCM for cross5 graph. Right: Cross entropy for edge probability between learned and ground-truth SCM for chain5",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 386,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 386,
+ 651
+ ],
+ "score": 1.0,
+ "content": "and cross5, comparing the learning process with prediction of edges to ground-truth intervention",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5
+ }
+ ],
+ "index": 7.25
+ }
+ ],
+ "page_idx": 16,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 288
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 288
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 288
+ ],
+ "score": 0.961,
+ "type": "image",
+ "image_path": "93a87e2e595b8857314a6cccac873b3eb129f202923fb425b9075b306cf5a5f5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 139,
+ 169,
+ 467,
+ 208.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 139,
+ 208.66666666666666,
+ 467,
+ 248.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 139,
+ 248.33333333333331,
+ 467,
+ 288.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 105,
+ 298,
+ 506,
+ 315
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "Figure 11: Cross entropy for edge probability between learned and ground-truth SCM. Left: The Earthquake dataset with 6 variables. Right:",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 203,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 203,
+ 316
+ ],
+ "score": 1.0,
+ "content": "The Asia dataset with 8 variables",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 624
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 624
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 624
+ ],
+ "score": 0.971,
+ "type": "image",
+ "image_path": "97ce72bf68cf2b4152045f11fbfda91a63a4d62a35f39a1aa1fc0b78630d6cf9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 150,
+ 480,
+ 469,
+ 528.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 150,
+ 528.0,
+ 469,
+ 576.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 150,
+ 576.0,
+ 469,
+ 624.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 103,
+ 633,
+ 505,
+ 651
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "Figure 12: Left: SCM for cross5 graph. Right: Cross entropy for edge probability between learned and ground-truth SCM for chain5",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 642,
+ 386,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 642,
+ 386,
+ 651
+ ],
+ "score": 1.0,
+ "content": "and cross5, comparing the learning process with prediction of edges to ground-truth intervention",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5
+ }
+ ],
+ "index": 7.25
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file
diff --git a/parse/train/H1gN6kSFwS/H1gN6kSFwS_model.json b/parse/train/H1gN6kSFwS/H1gN6kSFwS_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..5373a4fae5b80f588d57d56eeb0fa304e16cb98e
--- /dev/null
+++ b/parse/train/H1gN6kSFwS/H1gN6kSFwS_model.json
@@ -0,0 +1,25779 @@
+[
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 398,
+ 568,
+ 1304,
+ 568,
+ 1304,
+ 1177,
+ 398,
+ 1177
+ ],
+ "score": 0.984
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1668,
+ 1406,
+ 1668,
+ 1406,
+ 2033,
+ 298,
+ 2033
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1315,
+ 1405,
+ 1315,
+ 1405,
+ 1651,
+ 298,
+ 1651
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 392,
+ 219,
+ 1307,
+ 219,
+ 1307,
+ 323,
+ 392,
+ 323
+ ],
+ "score": 0.952
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 75,
+ 858,
+ 75,
+ 858,
+ 104,
+ 298,
+ 104
+ ],
+ "score": 0.896
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 302,
+ 1245,
+ 573,
+ 1245,
+ 573,
+ 1279,
+ 302,
+ 1279
+ ],
+ "score": 0.891
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 773,
+ 495,
+ 926,
+ 495,
+ 926,
+ 529,
+ 773,
+ 529
+ ],
+ "score": 0.852
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 856,
+ 2088,
+ 856,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.724
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 313,
+ 377,
+ 681,
+ 377,
+ 681,
+ 438,
+ 313,
+ 438
+ ],
+ "score": 0.583
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 316,
+ 378,
+ 555,
+ 378,
+ 555,
+ 406,
+ 316,
+ 406
+ ],
+ "score": 0.424
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 388.0,
+ 216.0,
+ 1313.0,
+ 216.0,
+ 1313.0,
+ 272.0,
+ 388.0,
+ 272.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 562.0,
+ 273.0,
+ 1139.0,
+ 273.0,
+ 1139.0,
+ 326.0,
+ 562.0,
+ 326.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 72.0,
+ 858.0,
+ 72.0,
+ 858.0,
+ 108.0,
+ 297.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1241.0,
+ 579.0,
+ 1241.0,
+ 579.0,
+ 1287.0,
+ 294.0,
+ 1287.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 769.0,
+ 494.0,
+ 932.0,
+ 494.0,
+ 932.0,
+ 532.0,
+ 769.0,
+ 532.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 840.0,
+ 2088.0,
+ 857.0,
+ 2088.0,
+ 857.0,
+ 2116.0,
+ 840.0,
+ 2116.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 310.0,
+ 371.0,
+ 561.0,
+ 371.0,
+ 561.0,
+ 415.0,
+ 310.0,
+ 415.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 395.0,
+ 570.0,
+ 1304.0,
+ 570.0,
+ 1304.0,
+ 602.0,
+ 395.0,
+ 602.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 599.0,
+ 1308.0,
+ 599.0,
+ 1308.0,
+ 633.0,
+ 393.0,
+ 633.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 630.0,
+ 1305.0,
+ 630.0,
+ 1305.0,
+ 665.0,
+ 393.0,
+ 665.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 660.0,
+ 1308.0,
+ 660.0,
+ 1308.0,
+ 696.0,
+ 394.0,
+ 696.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 690.0,
+ 1304.0,
+ 690.0,
+ 1304.0,
+ 723.0,
+ 394.0,
+ 723.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 722.0,
+ 1307.0,
+ 722.0,
+ 1307.0,
+ 754.0,
+ 394.0,
+ 754.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 395.0,
+ 752.0,
+ 1305.0,
+ 752.0,
+ 1305.0,
+ 785.0,
+ 395.0,
+ 785.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 395.0,
+ 783.0,
+ 1307.0,
+ 783.0,
+ 1307.0,
+ 816.0,
+ 395.0,
+ 816.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 813.0,
+ 1305.0,
+ 813.0,
+ 1305.0,
+ 844.0,
+ 394.0,
+ 844.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 843.0,
+ 1306.0,
+ 843.0,
+ 1306.0,
+ 876.0,
+ 394.0,
+ 876.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 874.0,
+ 1305.0,
+ 874.0,
+ 1305.0,
+ 906.0,
+ 394.0,
+ 906.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 391.0,
+ 899.0,
+ 1306.0,
+ 899.0,
+ 1306.0,
+ 942.0,
+ 391.0,
+ 942.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 934.0,
+ 1308.0,
+ 934.0,
+ 1308.0,
+ 968.0,
+ 393.0,
+ 968.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 395.0,
+ 967.0,
+ 1308.0,
+ 967.0,
+ 1308.0,
+ 997.0,
+ 395.0,
+ 997.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 994.0,
+ 1306.0,
+ 994.0,
+ 1306.0,
+ 1029.0,
+ 393.0,
+ 1029.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 1025.0,
+ 1305.0,
+ 1025.0,
+ 1305.0,
+ 1060.0,
+ 393.0,
+ 1060.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 1057.0,
+ 1305.0,
+ 1057.0,
+ 1305.0,
+ 1090.0,
+ 394.0,
+ 1090.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 1085.0,
+ 1305.0,
+ 1085.0,
+ 1305.0,
+ 1119.0,
+ 394.0,
+ 1119.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 1115.0,
+ 1304.0,
+ 1115.0,
+ 1304.0,
+ 1151.0,
+ 394.0,
+ 1151.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 394.0,
+ 1147.0,
+ 1272.0,
+ 1147.0,
+ 1272.0,
+ 1183.0,
+ 394.0,
+ 1183.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1669.0,
+ 1406.0,
+ 1669.0,
+ 1406.0,
+ 1703.0,
+ 295.0,
+ 1703.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1697.0,
+ 1406.0,
+ 1697.0,
+ 1406.0,
+ 1732.0,
+ 295.0,
+ 1732.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1728.0,
+ 1407.0,
+ 1728.0,
+ 1407.0,
+ 1761.0,
+ 293.0,
+ 1761.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1757.0,
+ 1409.0,
+ 1757.0,
+ 1409.0,
+ 1795.0,
+ 293.0,
+ 1795.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1789.0,
+ 1408.0,
+ 1789.0,
+ 1408.0,
+ 1824.0,
+ 293.0,
+ 1824.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1818.0,
+ 1407.0,
+ 1818.0,
+ 1407.0,
+ 1857.0,
+ 292.0,
+ 1857.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1852.0,
+ 1404.0,
+ 1852.0,
+ 1404.0,
+ 1886.0,
+ 295.0,
+ 1886.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1880.0,
+ 1407.0,
+ 1880.0,
+ 1407.0,
+ 1917.0,
+ 293.0,
+ 1917.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1910.0,
+ 1406.0,
+ 1910.0,
+ 1406.0,
+ 1946.0,
+ 293.0,
+ 1946.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1943.0,
+ 1406.0,
+ 1943.0,
+ 1406.0,
+ 1976.0,
+ 295.0,
+ 1976.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1973.0,
+ 1407.0,
+ 1973.0,
+ 1407.0,
+ 2007.0,
+ 295.0,
+ 2007.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 2003.0,
+ 1269.0,
+ 2003.0,
+ 1269.0,
+ 2037.0,
+ 295.0,
+ 2037.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1316.0,
+ 1407.0,
+ 1316.0,
+ 1407.0,
+ 1354.0,
+ 293.0,
+ 1354.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1348.0,
+ 1403.0,
+ 1348.0,
+ 1403.0,
+ 1381.0,
+ 295.0,
+ 1381.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1376.0,
+ 1405.0,
+ 1376.0,
+ 1405.0,
+ 1414.0,
+ 293.0,
+ 1414.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1405.0,
+ 1406.0,
+ 1405.0,
+ 1406.0,
+ 1445.0,
+ 292.0,
+ 1445.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1438.0,
+ 1407.0,
+ 1438.0,
+ 1407.0,
+ 1473.0,
+ 295.0,
+ 1473.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1467.0,
+ 1405.0,
+ 1467.0,
+ 1405.0,
+ 1504.0,
+ 293.0,
+ 1504.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1499.0,
+ 1405.0,
+ 1499.0,
+ 1405.0,
+ 1534.0,
+ 295.0,
+ 1534.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1528.0,
+ 1405.0,
+ 1528.0,
+ 1405.0,
+ 1565.0,
+ 293.0,
+ 1565.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1561.0,
+ 1403.0,
+ 1561.0,
+ 1403.0,
+ 1596.0,
+ 295.0,
+ 1596.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1590.0,
+ 1405.0,
+ 1590.0,
+ 1405.0,
+ 1625.0,
+ 293.0,
+ 1625.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1619.0,
+ 994.0,
+ 1619.0,
+ 994.0,
+ 1655.0,
+ 295.0,
+ 1655.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 315.0,
+ 377.0,
+ 560.0,
+ 377.0,
+ 560.0,
+ 410.0,
+ 315.0,
+ 410.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 311.0,
+ 407.0,
+ 682.0,
+ 407.0,
+ 682.0,
+ 440.0,
+ 311.0,
+ 440.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 0,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 519,
+ 1405,
+ 519,
+ 1405,
+ 1106,
+ 298,
+ 1106
+ ],
+ "score": 0.984
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1691,
+ 1404,
+ 1691,
+ 1404,
+ 1908,
+ 298,
+ 1908
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 229,
+ 1404,
+ 229,
+ 1404,
+ 505,
+ 298,
+ 505
+ ],
+ "score": 0.976
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 300,
+ 1943,
+ 1402,
+ 1943,
+ 1402,
+ 2034,
+ 300,
+ 2034
+ ],
+ "score": 0.971
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 369,
+ 1178,
+ 1404,
+ 1178,
+ 1404,
+ 1449,
+ 369,
+ 1449
+ ],
+ "score": 0.962
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1558,
+ 1400,
+ 1558,
+ 1400,
+ 1623,
+ 299,
+ 1623
+ ],
+ "score": 0.948
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 572,
+ 1637,
+ 1127,
+ 1637,
+ 1127,
+ 1679,
+ 572,
+ 1679
+ ],
+ "score": 0.935
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1121,
+ 1110,
+ 1121,
+ 1110,
+ 1153,
+ 298,
+ 1153
+ ],
+ "score": 0.909
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 1491,
+ 579,
+ 1491,
+ 579,
+ 1526,
+ 300,
+ 1526
+ ],
+ "score": 0.907
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 75,
+ 858,
+ 75,
+ 858,
+ 104,
+ 298,
+ 104
+ ],
+ "score": 0.896
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1366,
+ 1643,
+ 1399,
+ 1643,
+ 1399,
+ 1672,
+ 1366,
+ 1672
+ ],
+ "score": 0.866
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 858,
+ 2088,
+ 858,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.712
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 859,
+ 2088,
+ 859,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.126
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1027,
+ 1722,
+ 1218,
+ 1722,
+ 1218,
+ 1758,
+ 1027,
+ 1758
+ ],
+ "score": 0.94,
+ "latex": "C \\in \\{ 0 , 1 \\} ^ { M \\times M }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 780,
+ 1693,
+ 876,
+ 1693,
+ 876,
+ 1727,
+ 780,
+ 1727
+ ],
+ "score": 0.93,
+ "latex": "p a ( i , C )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 774,
+ 1756,
+ 941,
+ 1756,
+ 941,
+ 1791,
+ 774,
+ 1791
+ ],
+ "score": 0.92,
+ "latex": "X _ { j } \\in X _ { p a ( i , C ) }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1289,
+ 1725,
+ 1375,
+ 1725,
+ 1375,
+ 1758,
+ 1289,
+ 1758
+ ],
+ "score": 0.91,
+ "latex": "c _ { i j } = 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1098,
+ 1816,
+ 1186,
+ 1816,
+ 1186,
+ 1848,
+ 1098,
+ 1848
+ ],
+ "score": 0.9,
+ "latex": "\\mathrm { T r } ( C ^ { n } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1092,
+ 978,
+ 1143,
+ 978,
+ 1143,
+ 1011,
+ 1092,
+ 1011
+ ],
+ "score": 0.89,
+ "latex": "2 ^ { M ^ { 2 } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 995,
+ 1757,
+ 1031,
+ 1757,
+ 1031,
+ 1788,
+ 995,
+ 1788
+ ],
+ "score": 0.89,
+ "latex": "X _ { j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1973,
+ 389,
+ 1973,
+ 389,
+ 2001,
+ 298,
+ 2001
+ ],
+ "score": 0.89,
+ "latex": "M = 2"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 571,
+ 1638,
+ 1127,
+ 1638,
+ 1127,
+ 1679,
+ 571,
+ 1679
+ ],
+ "score": 0.89,
+ "latex": "X _ { i } : = f _ { i } ( X _ { p a ( i , C ) } , N _ { i } ) , \\quad \\forall i \\in \\{ 0 , \\dots , M - 1 \\}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 373,
+ 1695,
+ 406,
+ 1695,
+ 406,
+ 1724,
+ 373,
+ 1724
+ ],
+ "score": 0.88,
+ "latex": "N _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1596,
+ 331,
+ 1596,
+ 331,
+ 1622,
+ 298,
+ 1622
+ ],
+ "score": 0.87,
+ "latex": "X _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 569,
+ 2003,
+ 661,
+ 2003,
+ 661,
+ 2030,
+ 569,
+ 2030
+ ],
+ "score": 0.87,
+ "latex": "B A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 467,
+ 2003,
+ 558,
+ 2003,
+ 558,
+ 2031,
+ 467,
+ 2031
+ ],
+ "score": 0.87,
+ "latex": "A B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1246,
+ 1758,
+ 1279,
+ 1758,
+ 1279,
+ 1785,
+ 1246,
+ 1785
+ ],
+ "score": 0.87,
+ "latex": "X _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 486,
+ 1820,
+ 520,
+ 1820,
+ 520,
+ 1850,
+ 486,
+ 1850
+ ],
+ "score": 0.87,
+ "latex": "c _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 634,
+ 1786,
+ 672,
+ 1786,
+ 672,
+ 1812,
+ 634,
+ 1812
+ ],
+ "score": 0.86,
+ "latex": "C ^ { n }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1140,
+ 1563,
+ 1172,
+ 1563,
+ 1172,
+ 1588,
+ 1140,
+ 1588
+ ],
+ "score": 0.85,
+ "latex": "M"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1315,
+ 1787,
+ 1331,
+ 1787,
+ 1331,
+ 1817,
+ 1315,
+ 1817
+ ],
+ "score": 0.83,
+ "latex": "j"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 544,
+ 1726,
+ 569,
+ 1726,
+ 569,
+ 1752,
+ 544,
+ 1752
+ ],
+ "score": 0.83,
+ "latex": "C"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 476,
+ 1757,
+ 493,
+ 1757,
+ 493,
+ 1787,
+ 476,
+ 1787
+ ],
+ "score": 0.83,
+ "latex": "j"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 909,
+ 1975,
+ 931,
+ 1975,
+ 931,
+ 2000,
+ 909,
+ 2000
+ ],
+ "score": 0.83,
+ "latex": "A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 988,
+ 1978,
+ 1012,
+ 1978,
+ 1012,
+ 2000,
+ 988,
+ 2000
+ ],
+ "score": 0.83,
+ "latex": "B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 696,
+ 2005,
+ 763,
+ 2005,
+ 763,
+ 2030,
+ 696,
+ 2030
+ ],
+ "score": 0.81,
+ "latex": "A \\bot B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1316,
+ 1848,
+ 1340,
+ 1848,
+ 1340,
+ 1873,
+ 1316,
+ 1873
+ ],
+ "score": 0.81,
+ "latex": "C"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 632,
+ 950,
+ 649,
+ 950,
+ 649,
+ 980,
+ 632,
+ 980
+ ],
+ "score": 0.81,
+ "latex": "j"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 568,
+ 985,
+ 585,
+ 985,
+ 585,
+ 1015,
+ 568,
+ 1015
+ ],
+ "score": 0.81,
+ "latex": "j"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 999,
+ 1792,
+ 1017,
+ 1792,
+ 1017,
+ 1812,
+ 999,
+ 1812
+ ],
+ "score": 0.8,
+ "latex": "n"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1351,
+ 1761,
+ 1370,
+ 1761,
+ 1370,
+ 1782,
+ 1351,
+ 1782
+ ],
+ "score": 0.8,
+ "latex": "n"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 299,
+ 1730,
+ 309,
+ 1730,
+ 309,
+ 1752,
+ 299,
+ 1752
+ ],
+ "score": 0.79,
+ "latex": "i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1296,
+ 2004,
+ 1320,
+ 2004,
+ 1320,
+ 2030,
+ 1296,
+ 2030
+ ],
+ "score": 0.79,
+ "latex": "A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1211,
+ 1788,
+ 1223,
+ 1788,
+ 1223,
+ 1812,
+ 1211,
+ 1812
+ ],
+ "score": 0.79,
+ "latex": "i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 357,
+ 1757,
+ 369,
+ 1757,
+ 369,
+ 1782,
+ 357,
+ 1782
+ ],
+ "score": 0.78,
+ "latex": "i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1384,
+ 950,
+ 1401,
+ 950,
+ 1401,
+ 980,
+ 1384,
+ 980
+ ],
+ "score": 0.77,
+ "latex": "j"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 710,
+ 1821,
+ 729,
+ 1821,
+ 729,
+ 1843,
+ 710,
+ 1843
+ ],
+ "score": 0.75,
+ "latex": "n"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1144,
+ 950,
+ 1157,
+ 950,
+ 1157,
+ 975,
+ 1144,
+ 975
+ ],
+ "score": 0.75,
+ "latex": "i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 405,
+ 1851,
+ 425,
+ 1851,
+ 425,
+ 1874,
+ 405,
+ 1874
+ ],
+ "score": 0.73,
+ "latex": "\\mathbf { \\nabla } \\cdot n"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1273,
+ 919,
+ 1286,
+ 919,
+ 1286,
+ 944,
+ 1273,
+ 944
+ ],
+ "score": 0.73,
+ "latex": "i"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1487.0,
+ 583.0,
+ 1487.0,
+ 583.0,
+ 1533.0,
+ 291.0,
+ 1533.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 72.0,
+ 858.0,
+ 72.0,
+ 858.0,
+ 108.0,
+ 297.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2085.0,
+ 862.0,
+ 2085.0,
+ 862.0,
+ 2120.0,
+ 839.0,
+ 2120.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 838.0,
+ 2085.0,
+ 862.0,
+ 2085.0,
+ 862.0,
+ 2120.0,
+ 838.0,
+ 2120.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 522.0,
+ 1408.0,
+ 522.0,
+ 1408.0,
+ 557.0,
+ 296.0,
+ 557.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 551.0,
+ 1405.0,
+ 551.0,
+ 1405.0,
+ 588.0,
+ 291.0,
+ 588.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 583.0,
+ 1405.0,
+ 583.0,
+ 1405.0,
+ 618.0,
+ 295.0,
+ 618.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 610.0,
+ 1408.0,
+ 610.0,
+ 1408.0,
+ 648.0,
+ 292.0,
+ 648.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 641.0,
+ 1403.0,
+ 641.0,
+ 1403.0,
+ 676.0,
+ 295.0,
+ 676.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 674.0,
+ 1405.0,
+ 674.0,
+ 1405.0,
+ 709.0,
+ 293.0,
+ 709.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 704.0,
+ 1405.0,
+ 704.0,
+ 1405.0,
+ 739.0,
+ 295.0,
+ 739.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 733.0,
+ 1405.0,
+ 733.0,
+ 1405.0,
+ 771.0,
+ 292.0,
+ 771.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 762.0,
+ 1406.0,
+ 762.0,
+ 1406.0,
+ 804.0,
+ 292.0,
+ 804.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 796.0,
+ 1406.0,
+ 796.0,
+ 1406.0,
+ 830.0,
+ 292.0,
+ 830.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 825.0,
+ 1406.0,
+ 825.0,
+ 1406.0,
+ 860.0,
+ 295.0,
+ 860.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 857.0,
+ 1405.0,
+ 857.0,
+ 1405.0,
+ 892.0,
+ 295.0,
+ 892.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 886.0,
+ 1405.0,
+ 886.0,
+ 1405.0,
+ 921.0,
+ 293.0,
+ 921.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 919.0,
+ 1272.0,
+ 919.0,
+ 1272.0,
+ 949.0,
+ 293.0,
+ 949.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1287.0,
+ 919.0,
+ 1407.0,
+ 919.0,
+ 1407.0,
+ 949.0,
+ 1287.0,
+ 949.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 948.0,
+ 631.0,
+ 948.0,
+ 631.0,
+ 983.0,
+ 295.0,
+ 983.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 650.0,
+ 948.0,
+ 1143.0,
+ 948.0,
+ 1143.0,
+ 983.0,
+ 650.0,
+ 983.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1158.0,
+ 948.0,
+ 1383.0,
+ 948.0,
+ 1383.0,
+ 983.0,
+ 1158.0,
+ 983.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 289.0,
+ 979.0,
+ 567.0,
+ 979.0,
+ 567.0,
+ 1021.0,
+ 289.0,
+ 1021.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 586.0,
+ 979.0,
+ 1091.0,
+ 979.0,
+ 1091.0,
+ 1021.0,
+ 586.0,
+ 1021.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1144.0,
+ 979.0,
+ 1407.0,
+ 979.0,
+ 1407.0,
+ 1021.0,
+ 1144.0,
+ 1021.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1012.0,
+ 1408.0,
+ 1012.0,
+ 1408.0,
+ 1051.0,
+ 293.0,
+ 1051.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1045.0,
+ 1407.0,
+ 1045.0,
+ 1407.0,
+ 1080.0,
+ 295.0,
+ 1080.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1075.0,
+ 1148.0,
+ 1075.0,
+ 1148.0,
+ 1111.0,
+ 295.0,
+ 1111.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1692.0,
+ 372.0,
+ 1692.0,
+ 372.0,
+ 1730.0,
+ 294.0,
+ 1730.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 407.0,
+ 1692.0,
+ 779.0,
+ 1692.0,
+ 779.0,
+ 1730.0,
+ 407.0,
+ 1730.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 877.0,
+ 1692.0,
+ 1408.0,
+ 1692.0,
+ 1408.0,
+ 1730.0,
+ 877.0,
+ 1730.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1715.0,
+ 298.0,
+ 1715.0,
+ 298.0,
+ 1762.0,
+ 291.0,
+ 1762.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 310.0,
+ 1715.0,
+ 543.0,
+ 1715.0,
+ 543.0,
+ 1762.0,
+ 310.0,
+ 1762.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 570.0,
+ 1715.0,
+ 1026.0,
+ 1715.0,
+ 1026.0,
+ 1762.0,
+ 570.0,
+ 1762.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1219.0,
+ 1715.0,
+ 1288.0,
+ 1715.0,
+ 1288.0,
+ 1762.0,
+ 1219.0,
+ 1762.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1376.0,
+ 1715.0,
+ 1410.0,
+ 1715.0,
+ 1410.0,
+ 1762.0,
+ 1376.0,
+ 1762.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1751.0,
+ 356.0,
+ 1751.0,
+ 356.0,
+ 1793.0,
+ 291.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 370.0,
+ 1751.0,
+ 475.0,
+ 1751.0,
+ 475.0,
+ 1793.0,
+ 370.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 494.0,
+ 1751.0,
+ 773.0,
+ 1751.0,
+ 773.0,
+ 1793.0,
+ 494.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 942.0,
+ 1751.0,
+ 994.0,
+ 1751.0,
+ 994.0,
+ 1793.0,
+ 942.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1032.0,
+ 1751.0,
+ 1245.0,
+ 1751.0,
+ 1245.0,
+ 1793.0,
+ 1032.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1280.0,
+ 1751.0,
+ 1350.0,
+ 1751.0,
+ 1350.0,
+ 1793.0,
+ 1280.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1371.0,
+ 1751.0,
+ 1407.0,
+ 1751.0,
+ 1407.0,
+ 1793.0,
+ 1371.0,
+ 1793.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1783.0,
+ 633.0,
+ 1783.0,
+ 633.0,
+ 1819.0,
+ 292.0,
+ 1819.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 673.0,
+ 1783.0,
+ 998.0,
+ 1783.0,
+ 998.0,
+ 1819.0,
+ 673.0,
+ 1819.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1018.0,
+ 1783.0,
+ 1210.0,
+ 1783.0,
+ 1210.0,
+ 1819.0,
+ 1018.0,
+ 1819.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1224.0,
+ 1783.0,
+ 1314.0,
+ 1783.0,
+ 1314.0,
+ 1819.0,
+ 1224.0,
+ 1819.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1332.0,
+ 1783.0,
+ 1406.0,
+ 1783.0,
+ 1406.0,
+ 1819.0,
+ 1332.0,
+ 1819.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1814.0,
+ 485.0,
+ 1814.0,
+ 485.0,
+ 1850.0,
+ 292.0,
+ 1850.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 521.0,
+ 1814.0,
+ 709.0,
+ 1814.0,
+ 709.0,
+ 1850.0,
+ 521.0,
+ 1850.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 730.0,
+ 1814.0,
+ 1097.0,
+ 1814.0,
+ 1097.0,
+ 1850.0,
+ 730.0,
+ 1850.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1187.0,
+ 1814.0,
+ 1406.0,
+ 1814.0,
+ 1406.0,
+ 1850.0,
+ 1187.0,
+ 1850.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1845.0,
+ 404.0,
+ 1845.0,
+ 404.0,
+ 1880.0,
+ 294.0,
+ 1880.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 426.0,
+ 1845.0,
+ 1315.0,
+ 1845.0,
+ 1315.0,
+ 1880.0,
+ 426.0,
+ 1880.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1341.0,
+ 1845.0,
+ 1404.0,
+ 1845.0,
+ 1404.0,
+ 1880.0,
+ 1341.0,
+ 1880.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1878.0,
+ 749.0,
+ 1878.0,
+ 749.0,
+ 1908.0,
+ 296.0,
+ 1908.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 227.0,
+ 1405.0,
+ 227.0,
+ 1405.0,
+ 267.0,
+ 293.0,
+ 267.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 260.0,
+ 1404.0,
+ 260.0,
+ 1404.0,
+ 297.0,
+ 294.0,
+ 297.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 291.0,
+ 1406.0,
+ 291.0,
+ 1406.0,
+ 328.0,
+ 292.0,
+ 328.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 321.0,
+ 1406.0,
+ 321.0,
+ 1406.0,
+ 358.0,
+ 294.0,
+ 358.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 350.0,
+ 1409.0,
+ 350.0,
+ 1409.0,
+ 389.0,
+ 291.0,
+ 389.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 380.0,
+ 1409.0,
+ 380.0,
+ 1409.0,
+ 418.0,
+ 293.0,
+ 418.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 410.0,
+ 1406.0,
+ 410.0,
+ 1406.0,
+ 449.0,
+ 292.0,
+ 449.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 436.0,
+ 1406.0,
+ 436.0,
+ 1406.0,
+ 483.0,
+ 291.0,
+ 483.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 474.0,
+ 482.0,
+ 474.0,
+ 482.0,
+ 507.0,
+ 295.0,
+ 507.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1942.0,
+ 1404.0,
+ 1942.0,
+ 1404.0,
+ 1975.0,
+ 295.0,
+ 1975.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 390.0,
+ 1971.0,
+ 908.0,
+ 1971.0,
+ 908.0,
+ 2008.0,
+ 390.0,
+ 2008.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 932.0,
+ 1971.0,
+ 987.0,
+ 1971.0,
+ 987.0,
+ 2008.0,
+ 932.0,
+ 2008.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1013.0,
+ 1971.0,
+ 1404.0,
+ 1971.0,
+ 1404.0,
+ 2008.0,
+ 1013.0,
+ 2008.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1999.0,
+ 466.0,
+ 1999.0,
+ 466.0,
+ 2039.0,
+ 294.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 559.0,
+ 1999.0,
+ 568.0,
+ 1999.0,
+ 568.0,
+ 2039.0,
+ 559.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 662.0,
+ 1999.0,
+ 695.0,
+ 1999.0,
+ 695.0,
+ 2039.0,
+ 662.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 764.0,
+ 1999.0,
+ 1295.0,
+ 1999.0,
+ 1295.0,
+ 2039.0,
+ 764.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1321.0,
+ 1999.0,
+ 1406.0,
+ 1999.0,
+ 1406.0,
+ 2039.0,
+ 1321.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 369.0,
+ 1180.0,
+ 1404.0,
+ 1180.0,
+ 1404.0,
+ 1213.0,
+ 369.0,
+ 1213.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 1213.0,
+ 1404.0,
+ 1213.0,
+ 1404.0,
+ 1243.0,
+ 393.0,
+ 1243.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 393.0,
+ 1242.0,
+ 1406.0,
+ 1242.0,
+ 1406.0,
+ 1278.0,
+ 393.0,
+ 1278.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 396.0,
+ 1276.0,
+ 593.0,
+ 1276.0,
+ 593.0,
+ 1305.0,
+ 396.0,
+ 1305.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 365.0,
+ 1311.0,
+ 1407.0,
+ 1311.0,
+ 1407.0,
+ 1348.0,
+ 365.0,
+ 1348.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 395.0,
+ 1345.0,
+ 1306.0,
+ 1345.0,
+ 1306.0,
+ 1377.0,
+ 395.0,
+ 1377.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 366.0,
+ 1384.0,
+ 1404.0,
+ 1384.0,
+ 1404.0,
+ 1419.0,
+ 366.0,
+ 1419.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 395.0,
+ 1419.0,
+ 857.0,
+ 1419.0,
+ 857.0,
+ 1448.0,
+ 395.0,
+ 1448.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1558.0,
+ 1139.0,
+ 1558.0,
+ 1139.0,
+ 1594.0,
+ 295.0,
+ 1594.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1173.0,
+ 1558.0,
+ 1404.0,
+ 1558.0,
+ 1404.0,
+ 1594.0,
+ 1173.0,
+ 1594.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 332.0,
+ 1590.0,
+ 702.0,
+ 1590.0,
+ 702.0,
+ 1625.0,
+ 332.0,
+ 1625.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 298.0,
+ 1121.0,
+ 1113.0,
+ 1121.0,
+ 1113.0,
+ 1156.0,
+ 298.0,
+ 1156.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 1,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 571,
+ 1404,
+ 571,
+ 1404,
+ 967,
+ 297,
+ 967
+ ],
+ "score": 0.984
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1336,
+ 1404,
+ 1336,
+ 1404,
+ 1672,
+ 298,
+ 1672
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1766,
+ 1403,
+ 1766,
+ 1403,
+ 2035,
+ 298,
+ 2035
+ ],
+ "score": 0.982
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1074,
+ 1405,
+ 1074,
+ 1405,
+ 1258,
+ 298,
+ 1258
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 368,
+ 1404,
+ 368,
+ 1404,
+ 555,
+ 297,
+ 555
+ ],
+ "score": 0.978
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 228,
+ 1403,
+ 228,
+ 1403,
+ 353,
+ 299,
+ 353
+ ],
+ "score": 0.973
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 1279,
+ 600,
+ 1279,
+ 600,
+ 1310,
+ 300,
+ 1310
+ ],
+ "score": 0.895
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 1710,
+ 997,
+ 1710,
+ 997,
+ 1741,
+ 299,
+ 1741
+ ],
+ "score": 0.879
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 76,
+ 857,
+ 76,
+ 857,
+ 104,
+ 298,
+ 104
+ ],
+ "score": 0.869
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 1004,
+ 1337,
+ 1004,
+ 1337,
+ 1039,
+ 299,
+ 1039
+ ],
+ "score": 0.841
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 858,
+ 2088,
+ 858,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.605
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 859,
+ 2088,
+ 859,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.417
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 300,
+ 76,
+ 856,
+ 76,
+ 856,
+ 104,
+ 300,
+ 104
+ ],
+ "score": 0.104
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 908,
+ 459,
+ 1086,
+ 459,
+ 1086,
+ 498,
+ 908,
+ 498
+ ],
+ "score": 0.94,
+ "latex": "\\begin{array} { r } { \\sigma ( x ) \\ = \\ \\frac { \\bar { } 1 } { 1 + e ^ { - x } } } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1218,
+ 1828,
+ 1312,
+ 1828,
+ 1312,
+ 1867,
+ 1218,
+ 1867
+ ],
+ "score": 0.93,
+ "latex": "O ( 2 ^ { M ^ { 2 } } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 460,
+ 593,
+ 460,
+ 593,
+ 494,
+ 298,
+ 494
+ ],
+ "score": 0.92,
+ "latex": "P ( c _ { B A } = 1 ) = \\sigma ( \\gamma _ { B A } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 951,
+ 230,
+ 1208,
+ 230,
+ 1208,
+ 264,
+ 951,
+ 264
+ ],
+ "score": 0.92,
+ "latex": "B = f _ { \\theta _ { B } } ( c _ { B A } \\cdot A , \\epsilon _ { B } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1227,
+ 475,
+ 1227,
+ 475,
+ 1257,
+ 298,
+ 1257
+ ],
+ "score": 0.92,
+ "latex": "\\theta = \\theta _ { \\mathrm { s l o w } } + \\theta _ { \\mathrm { f a s t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 328,
+ 292,
+ 582,
+ 292,
+ 582,
+ 324,
+ 328,
+ 324
+ ],
+ "score": 0.92,
+ "latex": "A = f _ { \\theta _ { A } } ( c _ { A B } \\cdot B , \\epsilon _ { A } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 596,
+ 754,
+ 822,
+ 754,
+ 822,
+ 788,
+ 596,
+ 788
+ ],
+ "score": 0.91,
+ "latex": "i \\in \\{ 0 , \\ldots , M - 1 \\}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1079,
+ 323,
+ 1181,
+ 323,
+ 1181,
+ 353,
+ 1079,
+ 353
+ ],
+ "score": 0.91,
+ "latex": "c _ { A B } = 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1231,
+ 323,
+ 1333,
+ 323,
+ 1333,
+ 354,
+ 1231,
+ 354
+ ],
+ "score": 0.91,
+ "latex": "c _ { B A } = 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 658,
+ 495,
+ 701,
+ 495,
+ 701,
+ 528,
+ 658,
+ 528
+ ],
+ "score": 0.91,
+ "latex": "f _ { \\theta _ { B } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 262,
+ 408,
+ 262,
+ 408,
+ 292,
+ 298,
+ 292
+ ],
+ "score": 0.91,
+ "latex": "\\epsilon _ { B } \\sim N _ { B }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 783,
+ 1166,
+ 837,
+ 1166,
+ 837,
+ 1197,
+ 783,
+ 1197
+ ],
+ "score": 0.91,
+ "latex": "\\theta _ { \\mathrm { s l o w } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 563,
+ 496,
+ 605,
+ 496,
+ 605,
+ 528,
+ 563,
+ 528
+ ],
+ "score": 0.91,
+ "latex": "f _ { \\theta _ { A } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1295,
+ 232,
+ 1397,
+ 232,
+ 1397,
+ 262,
+ 1295,
+ 262
+ ],
+ "score": 0.9,
+ "latex": "c _ { B A } = 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1048,
+ 429,
+ 1349,
+ 429,
+ 1349,
+ 463,
+ 1048,
+ 463
+ ],
+ "score": 0.9,
+ "latex": "P ( c _ { A B } = 1 ) ^ { \\bar { } } = \\sigma ( \\gamma _ { A B } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 960,
+ 1903,
+ 1029,
+ 1903,
+ 1029,
+ 1933,
+ 960,
+ 1933
+ ],
+ "score": 0.9,
+ "latex": "i j"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1191,
+ 1197,
+ 1237,
+ 1197,
+ 1237,
+ 1228,
+ 1191,
+ 1228
+ ],
+ "score": 0.9,
+ "latex": "\\theta _ { \\mathrm { f a s t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 781,
+ 293,
+ 890,
+ 293,
+ 890,
+ 323,
+ 781,
+ 323
+ ],
+ "score": 0.89,
+ "latex": "\\epsilon _ { A } \\sim N _ { A }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 646,
+ 1968,
+ 695,
+ 1968,
+ 695,
+ 2000,
+ 646,
+ 2000
+ ],
+ "score": 0.89,
+ "latex": "2 ^ { M ^ { 2 } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 979,
+ 1864,
+ 1030,
+ 1864,
+ 1030,
+ 1898,
+ 979,
+ 1898
+ ],
+ "score": 0.89,
+ "latex": "2 ^ { M ^ { 2 } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 669,
+ 293,
+ 770,
+ 293,
+ 770,
+ 322,
+ 669,
+ 322
+ ],
+ "score": 0.88,
+ "latex": "c _ { A B } = 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 966,
+ 1934,
+ 1011,
+ 1934,
+ 1011,
+ 1962,
+ 966,
+ 1962
+ ],
+ "score": 0.88,
+ "latex": "M ^ { 2 }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 669,
+ 2006,
+ 704,
+ 2006,
+ 704,
+ 2038,
+ 669,
+ 2038
+ ],
+ "score": 0.88,
+ "latex": "c _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1151,
+ 1938,
+ 1188,
+ 1938,
+ 1188,
+ 1969,
+ 1151,
+ 1969
+ ],
+ "score": 0.88,
+ "latex": "\\gamma _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1062,
+ 847,
+ 1088,
+ 847,
+ 1088,
+ 878,
+ 1062,
+ 878
+ ],
+ "score": 0.88,
+ "latex": "f _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 827,
+ 879,
+ 862,
+ 879,
+ 862,
+ 910,
+ 827,
+ 910
+ ],
+ "score": 0.87,
+ "latex": "c _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1250,
+ 1975,
+ 1284,
+ 1975,
+ 1284,
+ 2006,
+ 1250,
+ 2006
+ ],
+ "score": 0.87,
+ "latex": "c _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 385,
+ 404,
+ 436,
+ 404,
+ 436,
+ 431,
+ 385,
+ 431
+ ],
+ "score": 0.87,
+ "latex": "c _ { B A }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 374,
+ 1936,
+ 409,
+ 1936,
+ 409,
+ 1969,
+ 374,
+ 1969
+ ],
+ "score": 0.86,
+ "latex": "c _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 406,
+ 347,
+ 406,
+ 347,
+ 430,
+ 298,
+ 430
+ ],
+ "score": 0.86,
+ "latex": "c _ { A B }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 634,
+ 1906,
+ 668,
+ 1906,
+ 668,
+ 1936,
+ 634,
+ 1936
+ ],
+ "score": 0.86,
+ "latex": "c _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 546,
+ 755,
+ 581,
+ 755,
+ 581,
+ 785,
+ 546,
+ 785
+ ],
+ "score": 0.84,
+ "latex": "X _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1003,
+ 637,
+ 1024,
+ 637,
+ 1024,
+ 664,
+ 1003,
+ 664
+ ],
+ "score": 0.83,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 891,
+ 728,
+ 912,
+ 728,
+ 912,
+ 756,
+ 891,
+ 756
+ ],
+ "score": 0.82,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 404,
+ 1171,
+ 426,
+ 1171,
+ 426,
+ 1198,
+ 404,
+ 1198
+ ],
+ "score": 0.81,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 372,
+ 325,
+ 398,
+ 325,
+ 398,
+ 349,
+ 372,
+ 349
+ ],
+ "score": 0.81,
+ "latex": "B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 669,
+ 465,
+ 691,
+ 465,
+ 691,
+ 493,
+ 669,
+ 493
+ ],
+ "score": 0.81,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1286,
+ 577,
+ 1308,
+ 577,
+ 1308,
+ 604,
+ 1286,
+ 604
+ ],
+ "score": 0.81,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 824,
+ 725,
+ 841,
+ 725,
+ 841,
+ 751,
+ 824,
+ 751
+ ],
+ "score": 0.8,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 446,
+ 728,
+ 466,
+ 728,
+ 466,
+ 755,
+ 446,
+ 755
+ ],
+ "score": 0.8,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 913,
+ 262,
+ 939,
+ 262,
+ 939,
+ 289,
+ 913,
+ 289
+ ],
+ "score": 0.8,
+ "latex": "B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 232,
+ 322,
+ 232,
+ 322,
+ 258,
+ 298,
+ 258
+ ],
+ "score": 0.79,
+ "latex": "B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 674,
+ 232,
+ 700,
+ 232,
+ 700,
+ 259,
+ 674,
+ 259
+ ],
+ "score": 0.79,
+ "latex": "B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 323,
+ 322,
+ 323,
+ 322,
+ 349,
+ 298,
+ 349
+ ],
+ "score": 0.79,
+ "latex": "A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1368,
+ 848,
+ 1401,
+ 848,
+ 1401,
+ 873,
+ 1368,
+ 873
+ ],
+ "score": 0.78,
+ "latex": "M"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 432,
+ 323,
+ 432,
+ 323,
+ 457,
+ 298,
+ 457
+ ],
+ "score": 0.78,
+ "latex": "B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1021,
+ 264,
+ 1044,
+ 264,
+ 1044,
+ 289,
+ 1021,
+ 289
+ ],
+ "score": 0.78,
+ "latex": "A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 615,
+ 604,
+ 632,
+ 604,
+ 632,
+ 630,
+ 615,
+ 630
+ ],
+ "score": 0.77,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 398,
+ 786,
+ 422,
+ 786,
+ 422,
+ 812,
+ 398,
+ 812
+ ],
+ "score": 0.77,
+ "latex": "C"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 971,
+ 604,
+ 988,
+ 604,
+ 988,
+ 630,
+ 971,
+ 630
+ ],
+ "score": 0.76,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1328,
+ 401,
+ 1352,
+ 401,
+ 1352,
+ 427,
+ 1328,
+ 427
+ ],
+ "score": 0.76,
+ "latex": "A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1201,
+ 725,
+ 1234,
+ 725,
+ 1234,
+ 752,
+ 1201,
+ 752
+ ],
+ "score": 0.76,
+ "latex": "M"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 408,
+ 1398,
+ 441,
+ 1398,
+ 441,
+ 1426,
+ 408,
+ 1426
+ ],
+ "score": 0.73,
+ "latex": "M"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 599,
+ 232,
+ 623,
+ 232,
+ 623,
+ 258,
+ 599,
+ 258
+ ],
+ "score": 0.73,
+ "latex": "A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 570,
+ 1798,
+ 604,
+ 1798,
+ 604,
+ 1826,
+ 570,
+ 1826
+ ],
+ "score": 0.69,
+ "latex": "M"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 418,
+ 1521,
+ 451,
+ 1521,
+ 451,
+ 1548,
+ 418,
+ 1548
+ ],
+ "score": 0.67,
+ "latex": "M"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1279.0,
+ 604.0,
+ 1279.0,
+ 604.0,
+ 1312.0,
+ 296.0,
+ 1312.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1709.0,
+ 1000.0,
+ 1709.0,
+ 1000.0,
+ 1745.0,
+ 295.0,
+ 1745.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 298.0,
+ 73.0,
+ 858.0,
+ 73.0,
+ 858.0,
+ 108.0,
+ 298.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 289.0,
+ 1000.0,
+ 1341.0,
+ 1000.0,
+ 1341.0,
+ 1045.0,
+ 289.0,
+ 1045.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2085.0,
+ 860.0,
+ 2085.0,
+ 860.0,
+ 2116.0,
+ 839.0,
+ 2116.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 838.0,
+ 2085.0,
+ 862.0,
+ 2085.0,
+ 862.0,
+ 2118.0,
+ 838.0,
+ 2118.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 73.0,
+ 859.0,
+ 73.0,
+ 859.0,
+ 108.0,
+ 297.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 573.0,
+ 1285.0,
+ 573.0,
+ 1285.0,
+ 605.0,
+ 296.0,
+ 605.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1309.0,
+ 573.0,
+ 1404.0,
+ 573.0,
+ 1404.0,
+ 605.0,
+ 1309.0,
+ 605.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 604.0,
+ 614.0,
+ 604.0,
+ 614.0,
+ 636.0,
+ 295.0,
+ 636.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 633.0,
+ 604.0,
+ 970.0,
+ 604.0,
+ 970.0,
+ 636.0,
+ 633.0,
+ 636.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 989.0,
+ 604.0,
+ 1405.0,
+ 604.0,
+ 1405.0,
+ 636.0,
+ 989.0,
+ 636.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 632.0,
+ 1002.0,
+ 632.0,
+ 1002.0,
+ 668.0,
+ 294.0,
+ 668.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1025.0,
+ 632.0,
+ 1405.0,
+ 632.0,
+ 1405.0,
+ 668.0,
+ 1025.0,
+ 668.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 659.0,
+ 1409.0,
+ 659.0,
+ 1409.0,
+ 701.0,
+ 291.0,
+ 701.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 689.0,
+ 1406.0,
+ 689.0,
+ 1406.0,
+ 733.0,
+ 291.0,
+ 733.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 723.0,
+ 445.0,
+ 723.0,
+ 445.0,
+ 757.0,
+ 294.0,
+ 757.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 467.0,
+ 723.0,
+ 823.0,
+ 723.0,
+ 823.0,
+ 757.0,
+ 467.0,
+ 757.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 842.0,
+ 723.0,
+ 890.0,
+ 723.0,
+ 890.0,
+ 757.0,
+ 842.0,
+ 757.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 913.0,
+ 723.0,
+ 1200.0,
+ 723.0,
+ 1200.0,
+ 757.0,
+ 913.0,
+ 757.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1235.0,
+ 723.0,
+ 1405.0,
+ 723.0,
+ 1405.0,
+ 757.0,
+ 1235.0,
+ 757.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 750.0,
+ 545.0,
+ 750.0,
+ 545.0,
+ 795.0,
+ 292.0,
+ 795.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 582.0,
+ 750.0,
+ 595.0,
+ 750.0,
+ 595.0,
+ 795.0,
+ 582.0,
+ 795.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 823.0,
+ 750.0,
+ 1405.0,
+ 750.0,
+ 1405.0,
+ 795.0,
+ 823.0,
+ 795.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 785.0,
+ 397.0,
+ 785.0,
+ 397.0,
+ 821.0,
+ 295.0,
+ 821.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 423.0,
+ 785.0,
+ 1405.0,
+ 785.0,
+ 1405.0,
+ 821.0,
+ 423.0,
+ 821.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 814.0,
+ 1405.0,
+ 814.0,
+ 1405.0,
+ 854.0,
+ 294.0,
+ 854.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 843.0,
+ 1061.0,
+ 843.0,
+ 1061.0,
+ 881.0,
+ 292.0,
+ 881.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1089.0,
+ 843.0,
+ 1367.0,
+ 843.0,
+ 1367.0,
+ 881.0,
+ 1089.0,
+ 881.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1402.0,
+ 843.0,
+ 1405.0,
+ 843.0,
+ 1405.0,
+ 881.0,
+ 1402.0,
+ 881.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 874.0,
+ 826.0,
+ 874.0,
+ 826.0,
+ 914.0,
+ 294.0,
+ 914.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 863.0,
+ 874.0,
+ 1406.0,
+ 874.0,
+ 1406.0,
+ 914.0,
+ 863.0,
+ 914.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 906.0,
+ 1406.0,
+ 906.0,
+ 1406.0,
+ 941.0,
+ 292.0,
+ 941.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 937.0,
+ 467.0,
+ 937.0,
+ 467.0,
+ 970.0,
+ 293.0,
+ 970.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1337.0,
+ 1405.0,
+ 1337.0,
+ 1405.0,
+ 1371.0,
+ 296.0,
+ 1371.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1368.0,
+ 1404.0,
+ 1368.0,
+ 1404.0,
+ 1402.0,
+ 294.0,
+ 1402.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1396.0,
+ 407.0,
+ 1396.0,
+ 407.0,
+ 1432.0,
+ 293.0,
+ 1432.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 442.0,
+ 1396.0,
+ 1405.0,
+ 1396.0,
+ 1405.0,
+ 1432.0,
+ 442.0,
+ 1432.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1427.0,
+ 1406.0,
+ 1427.0,
+ 1406.0,
+ 1463.0,
+ 292.0,
+ 1463.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1458.0,
+ 1404.0,
+ 1458.0,
+ 1404.0,
+ 1494.0,
+ 293.0,
+ 1494.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1489.0,
+ 1406.0,
+ 1489.0,
+ 1406.0,
+ 1524.0,
+ 294.0,
+ 1524.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1520.0,
+ 417.0,
+ 1520.0,
+ 417.0,
+ 1554.0,
+ 293.0,
+ 1554.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 452.0,
+ 1520.0,
+ 1406.0,
+ 1520.0,
+ 1406.0,
+ 1554.0,
+ 452.0,
+ 1554.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1550.0,
+ 1404.0,
+ 1550.0,
+ 1404.0,
+ 1585.0,
+ 294.0,
+ 1585.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1580.0,
+ 1404.0,
+ 1580.0,
+ 1404.0,
+ 1615.0,
+ 293.0,
+ 1615.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1609.0,
+ 1405.0,
+ 1609.0,
+ 1405.0,
+ 1646.0,
+ 293.0,
+ 1646.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1642.0,
+ 1245.0,
+ 1642.0,
+ 1245.0,
+ 1675.0,
+ 294.0,
+ 1675.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1765.0,
+ 1404.0,
+ 1765.0,
+ 1404.0,
+ 1803.0,
+ 293.0,
+ 1803.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1795.0,
+ 569.0,
+ 1795.0,
+ 569.0,
+ 1834.0,
+ 293.0,
+ 1834.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 605.0,
+ 1795.0,
+ 1405.0,
+ 1795.0,
+ 1405.0,
+ 1834.0,
+ 605.0,
+ 1834.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 289.0,
+ 1826.0,
+ 1217.0,
+ 1826.0,
+ 1217.0,
+ 1872.0,
+ 289.0,
+ 1872.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1313.0,
+ 1826.0,
+ 1409.0,
+ 1826.0,
+ 1409.0,
+ 1872.0,
+ 1313.0,
+ 1872.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 289.0,
+ 1865.0,
+ 978.0,
+ 1865.0,
+ 978.0,
+ 1908.0,
+ 289.0,
+ 1908.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1031.0,
+ 1865.0,
+ 1405.0,
+ 1865.0,
+ 1405.0,
+ 1908.0,
+ 1031.0,
+ 1908.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1898.0,
+ 633.0,
+ 1898.0,
+ 633.0,
+ 1940.0,
+ 292.0,
+ 1940.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 669.0,
+ 1898.0,
+ 959.0,
+ 1898.0,
+ 959.0,
+ 1940.0,
+ 669.0,
+ 1940.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1030.0,
+ 1898.0,
+ 1407.0,
+ 1898.0,
+ 1407.0,
+ 1940.0,
+ 1030.0,
+ 1940.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1933.0,
+ 373.0,
+ 1933.0,
+ 373.0,
+ 1972.0,
+ 293.0,
+ 1972.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 410.0,
+ 1933.0,
+ 965.0,
+ 1933.0,
+ 965.0,
+ 1972.0,
+ 410.0,
+ 1972.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1012.0,
+ 1933.0,
+ 1150.0,
+ 1933.0,
+ 1150.0,
+ 1972.0,
+ 1012.0,
+ 1972.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1189.0,
+ 1933.0,
+ 1405.0,
+ 1933.0,
+ 1405.0,
+ 1972.0,
+ 1189.0,
+ 1972.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 289.0,
+ 1961.0,
+ 645.0,
+ 1961.0,
+ 645.0,
+ 2013.0,
+ 289.0,
+ 2013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 696.0,
+ 1961.0,
+ 1249.0,
+ 1961.0,
+ 1249.0,
+ 2013.0,
+ 696.0,
+ 2013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1285.0,
+ 1961.0,
+ 1412.0,
+ 1961.0,
+ 1412.0,
+ 2013.0,
+ 1285.0,
+ 2013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 2003.0,
+ 668.0,
+ 2003.0,
+ 668.0,
+ 2039.0,
+ 294.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 705.0,
+ 2003.0,
+ 1405.0,
+ 2003.0,
+ 1405.0,
+ 2039.0,
+ 705.0,
+ 2039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1071.0,
+ 1409.0,
+ 1071.0,
+ 1409.0,
+ 1113.0,
+ 293.0,
+ 1113.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1103.0,
+ 1406.0,
+ 1103.0,
+ 1406.0,
+ 1138.0,
+ 295.0,
+ 1138.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1134.0,
+ 1405.0,
+ 1134.0,
+ 1405.0,
+ 1171.0,
+ 293.0,
+ 1171.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1166.0,
+ 403.0,
+ 1166.0,
+ 403.0,
+ 1202.0,
+ 295.0,
+ 1202.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 427.0,
+ 1166.0,
+ 782.0,
+ 1166.0,
+ 782.0,
+ 1202.0,
+ 427.0,
+ 1202.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 838.0,
+ 1166.0,
+ 1405.0,
+ 1166.0,
+ 1405.0,
+ 1202.0,
+ 838.0,
+ 1202.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1194.0,
+ 1190.0,
+ 1194.0,
+ 1190.0,
+ 1233.0,
+ 293.0,
+ 1233.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1238.0,
+ 1194.0,
+ 1406.0,
+ 1194.0,
+ 1406.0,
+ 1233.0,
+ 1238.0,
+ 1233.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1224.0,
+ 297.0,
+ 1224.0,
+ 297.0,
+ 1262.0,
+ 292.0,
+ 1262.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 476.0,
+ 1224.0,
+ 1408.0,
+ 1224.0,
+ 1408.0,
+ 1262.0,
+ 476.0,
+ 1262.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 367.0,
+ 1406.0,
+ 367.0,
+ 1406.0,
+ 401.0,
+ 294.0,
+ 401.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 397.0,
+ 297.0,
+ 397.0,
+ 297.0,
+ 435.0,
+ 291.0,
+ 435.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 348.0,
+ 397.0,
+ 384.0,
+ 397.0,
+ 384.0,
+ 435.0,
+ 348.0,
+ 435.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 437.0,
+ 397.0,
+ 1327.0,
+ 397.0,
+ 1327.0,
+ 435.0,
+ 437.0,
+ 435.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1353.0,
+ 397.0,
+ 1407.0,
+ 397.0,
+ 1407.0,
+ 435.0,
+ 1353.0,
+ 435.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 426.0,
+ 297.0,
+ 426.0,
+ 297.0,
+ 465.0,
+ 294.0,
+ 465.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 324.0,
+ 426.0,
+ 1047.0,
+ 426.0,
+ 1047.0,
+ 465.0,
+ 324.0,
+ 465.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1350.0,
+ 426.0,
+ 1406.0,
+ 426.0,
+ 1406.0,
+ 465.0,
+ 1350.0,
+ 465.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 287.0,
+ 451.0,
+ 297.0,
+ 451.0,
+ 297.0,
+ 509.0,
+ 287.0,
+ 509.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 594.0,
+ 451.0,
+ 668.0,
+ 451.0,
+ 668.0,
+ 509.0,
+ 594.0,
+ 509.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 692.0,
+ 451.0,
+ 907.0,
+ 451.0,
+ 907.0,
+ 509.0,
+ 692.0,
+ 509.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1087.0,
+ 451.0,
+ 1410.0,
+ 451.0,
+ 1410.0,
+ 509.0,
+ 1087.0,
+ 509.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 492.0,
+ 562.0,
+ 492.0,
+ 562.0,
+ 532.0,
+ 294.0,
+ 532.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 606.0,
+ 492.0,
+ 657.0,
+ 492.0,
+ 657.0,
+ 532.0,
+ 606.0,
+ 532.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 702.0,
+ 492.0,
+ 1405.0,
+ 492.0,
+ 1405.0,
+ 532.0,
+ 702.0,
+ 532.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 526.0,
+ 604.0,
+ 526.0,
+ 604.0,
+ 558.0,
+ 295.0,
+ 558.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 226.0,
+ 297.0,
+ 226.0,
+ 297.0,
+ 268.0,
+ 293.0,
+ 268.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 323.0,
+ 226.0,
+ 598.0,
+ 226.0,
+ 598.0,
+ 268.0,
+ 323.0,
+ 268.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 624.0,
+ 226.0,
+ 673.0,
+ 226.0,
+ 673.0,
+ 268.0,
+ 624.0,
+ 268.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 701.0,
+ 226.0,
+ 950.0,
+ 226.0,
+ 950.0,
+ 268.0,
+ 701.0,
+ 268.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1209.0,
+ 226.0,
+ 1294.0,
+ 226.0,
+ 1294.0,
+ 268.0,
+ 1209.0,
+ 268.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1398.0,
+ 226.0,
+ 1409.0,
+ 226.0,
+ 1409.0,
+ 268.0,
+ 1398.0,
+ 268.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 262.0,
+ 297.0,
+ 262.0,
+ 297.0,
+ 294.0,
+ 294.0,
+ 294.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 409.0,
+ 262.0,
+ 912.0,
+ 262.0,
+ 912.0,
+ 294.0,
+ 409.0,
+ 294.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 940.0,
+ 262.0,
+ 1020.0,
+ 262.0,
+ 1020.0,
+ 294.0,
+ 940.0,
+ 294.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1045.0,
+ 262.0,
+ 1405.0,
+ 262.0,
+ 1405.0,
+ 294.0,
+ 1045.0,
+ 294.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 292.0,
+ 327.0,
+ 292.0,
+ 327.0,
+ 324.0,
+ 293.0,
+ 324.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 583.0,
+ 292.0,
+ 668.0,
+ 292.0,
+ 668.0,
+ 324.0,
+ 583.0,
+ 324.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 771.0,
+ 292.0,
+ 780.0,
+ 292.0,
+ 780.0,
+ 324.0,
+ 771.0,
+ 324.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 891.0,
+ 292.0,
+ 1407.0,
+ 292.0,
+ 1407.0,
+ 324.0,
+ 891.0,
+ 324.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 321.0,
+ 297.0,
+ 321.0,
+ 297.0,
+ 357.0,
+ 294.0,
+ 357.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 323.0,
+ 321.0,
+ 371.0,
+ 321.0,
+ 371.0,
+ 357.0,
+ 323.0,
+ 357.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 399.0,
+ 321.0,
+ 1078.0,
+ 321.0,
+ 1078.0,
+ 357.0,
+ 399.0,
+ 357.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1182.0,
+ 321.0,
+ 1230.0,
+ 321.0,
+ 1230.0,
+ 357.0,
+ 1182.0,
+ 357.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1334.0,
+ 321.0,
+ 1344.0,
+ 321.0,
+ 1344.0,
+ 357.0,
+ 1334.0,
+ 357.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 2,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1049,
+ 1405,
+ 1049,
+ 1405,
+ 1327,
+ 297,
+ 1327
+ ],
+ "score": 0.982
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1340,
+ 1404,
+ 1340,
+ 1404,
+ 1494,
+ 298,
+ 1494
+ ],
+ "score": 0.979
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 751,
+ 371,
+ 1384,
+ 371,
+ 1384,
+ 728,
+ 751,
+ 728
+ ],
+ "score": 0.966
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 331,
+ 717,
+ 331,
+ 717,
+ 915,
+ 298,
+ 915
+ ],
+ "score": 0.964
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 229,
+ 1401,
+ 229,
+ 1401,
+ 323,
+ 299,
+ 323
+ ],
+ "score": 0.962
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1849,
+ 1403,
+ 1849,
+ 1403,
+ 1944,
+ 299,
+ 1944
+ ],
+ "score": 0.959
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 293,
+ 1605,
+ 1403,
+ 1605,
+ 1403,
+ 1670,
+ 293,
+ 1670
+ ],
+ "score": 0.955
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 561,
+ 1516,
+ 1137,
+ 1516,
+ 1137,
+ 1586,
+ 561,
+ 1586
+ ],
+ "score": 0.954
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 300,
+ 1747,
+ 1402,
+ 1747,
+ 1402,
+ 1815,
+ 300,
+ 1815
+ ],
+ "score": 0.952
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 736,
+ 762,
+ 1400,
+ 762,
+ 1400,
+ 878,
+ 736,
+ 878
+ ],
+ "score": 0.952
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 294,
+ 1977,
+ 1404,
+ 1977,
+ 1404,
+ 2033,
+ 294,
+ 2033
+ ],
+ "score": 0.942
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 657,
+ 1689,
+ 1041,
+ 1689,
+ 1041,
+ 1732,
+ 657,
+ 1732
+ ],
+ "score": 0.934
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 991,
+ 847,
+ 991,
+ 847,
+ 1022,
+ 299,
+ 1022
+ ],
+ "score": 0.9
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1366,
+ 1526,
+ 1399,
+ 1526,
+ 1399,
+ 1556,
+ 1366,
+ 1556
+ ],
+ "score": 0.885
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 75,
+ 858,
+ 75,
+ 858,
+ 104,
+ 298,
+ 104
+ ],
+ "score": 0.882
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1366,
+ 1694,
+ 1399,
+ 1694,
+ 1399,
+ 1724,
+ 1366,
+ 1724
+ ],
+ "score": 0.879
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 917,
+ 1071,
+ 917,
+ 1071,
+ 946,
+ 299,
+ 946
+ ],
+ "score": 0.798
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 840,
+ 2089,
+ 857,
+ 2089,
+ 857,
+ 2111,
+ 840,
+ 2111
+ ],
+ "score": 0.785
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1201,
+ 1141,
+ 1398,
+ 1141,
+ 1398,
+ 1177,
+ 1201,
+ 1177
+ ],
+ "score": 0.94,
+ "latex": "P _ { i } ( X _ { i } | X _ { p a ( i ) } ; \\theta _ { i } )"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 563,
+ 1514,
+ 1137,
+ 1514,
+ 1137,
+ 1588,
+ 563,
+ 1588
+ ],
+ "score": 0.93,
+ "latex": "\\mathcal { R } = - \\mathbb { E } _ { X \\sim D _ { \\mathrm { i n t } } } [ \\log \\mathbb { E } _ { C \\sim \\mathrm { B e r } ( \\gamma ) } [ \\prod _ { i } \\mathcal { L } _ { C , i } ( X ; \\theta _ { \\mathrm { s l o w } } ) ] ]"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 457,
+ 1142,
+ 633,
+ 1142,
+ 633,
+ 1173,
+ 457,
+ 1173
+ ],
+ "score": 0.93,
+ "latex": "\\theta = \\theta _ { \\mathrm { s l o w } } + \\theta _ { \\mathrm { f a s t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 718,
+ 1781,
+ 815,
+ 1781,
+ 815,
+ 1815,
+ 718,
+ 1815
+ ],
+ "score": 0.92,
+ "latex": "X _ { p a ( i , C ) }"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 657,
+ 1690,
+ 1042,
+ 1690,
+ 1042,
+ 1731,
+ 657,
+ 1731
+ ],
+ "score": 0.92,
+ "latex": "\\mathcal { L } _ { C , i } ( X ) = P ( X _ { i } | X _ { p a ( i , C ) } ; \\theta _ { \\mathrm { s l o w } } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 382,
+ 263,
+ 499,
+ 263,
+ 499,
+ 296,
+ 382,
+ 296
+ ],
+ "score": 0.91,
+ "latex": "c _ { i j } c _ { j i } = 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 629,
+ 425,
+ 687,
+ 425,
+ 687,
+ 462,
+ 629,
+ 462
+ ],
+ "score": 0.91,
+ "latex": "f _ { \\theta _ { i } } ( \\dot { ) }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 978,
+ 1112,
+ 1032,
+ 1112,
+ 1032,
+ 1142,
+ 978,
+ 1142
+ ],
+ "score": 0.91,
+ "latex": "\\theta _ { \\mathrm { s l o w } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1278,
+ 1433,
+ 1331,
+ 1433,
+ 1331,
+ 1464,
+ 1278,
+ 1464
+ ],
+ "score": 0.9,
+ "latex": "D _ { \\mathrm { i n t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1212,
+ 1113,
+ 1257,
+ 1113,
+ 1257,
+ 1142,
+ 1212,
+ 1142
+ ],
+ "score": 0.9,
+ "latex": "\\theta _ { \\mathrm { f a s t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 297,
+ 428,
+ 545,
+ 428,
+ 545,
+ 461,
+ 297,
+ 461
+ ],
+ "score": 0.9,
+ "latex": "X _ { 1 } , . . . , c _ { i m } \\times X _ { m } , \\epsilon _ { i } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1066,
+ 784,
+ 1141,
+ 784,
+ 1141,
+ 811,
+ 1066,
+ 811
+ ],
+ "score": 0.9,
+ "latex": "{ \\hat { A } } , { \\hat { B } } , { \\hat { C } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 783,
+ 833,
+ 879,
+ 833,
+ 879,
+ 857,
+ 783,
+ 857
+ ],
+ "score": 0.89,
+ "latex": "\\mathrm { B e r } ( \\sigma ( \\gamma ) )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 427,
+ 490,
+ 708,
+ 490,
+ 708,
+ 524,
+ 427,
+ 524
+ ],
+ "score": 0.89,
+ "latex": "c _ { i j } \\sim \\mathrm { B i n } ( \\mathrm { s i g m o i d } ( \\gamma _ { i j } ) )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 697,
+ 1750,
+ 730,
+ 1750,
+ 730,
+ 1779,
+ 697,
+ 1779
+ ],
+ "score": 0.89,
+ "latex": "X _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 398,
+ 391,
+ 714,
+ 391,
+ 714,
+ 425,
+ 398,
+ 425
+ ],
+ "score": 0.88,
+ "latex": "X _ { i } : = f _ { \\theta _ { i } } ( \\bar { c } _ { i 0 } \\times X _ { 0 } , c _ { i 1 } \\times"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1046,
+ 764,
+ 1112,
+ 764,
+ 1112,
+ 784,
+ 1046,
+ 784
+ ],
+ "score": 0.87,
+ "latex": "M = 3"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 344,
+ 491,
+ 369,
+ 491,
+ 369,
+ 520,
+ 344,
+ 520
+ ],
+ "score": 0.86,
+ "latex": "\\theta _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1121,
+ 763,
+ 1185,
+ 763,
+ 1185,
+ 784,
+ 1121,
+ 784
+ ],
+ "score": 0.86,
+ "latex": "N = 2"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 495,
+ 1404,
+ 591,
+ 1404,
+ 591,
+ 1435,
+ 495,
+ 1435
+ ],
+ "score": 0.86,
+ "latex": "( \\theta _ { \\mathrm { { s l o w } } } , \\gamma )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1052,
+ 1609,
+ 1104,
+ 1609,
+ 1104,
+ 1638,
+ 1052,
+ 1638
+ ],
+ "score": 0.86,
+ "latex": "D _ { \\mathrm { i n t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 892,
+ 1751,
+ 919,
+ 1751,
+ 919,
+ 1776,
+ 892,
+ 1776
+ ],
+ "score": 0.85,
+ "latex": "X"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 561,
+ 1373,
+ 618,
+ 1373,
+ 618,
+ 1403,
+ 561,
+ 1403
+ ],
+ "score": 0.84,
+ "latex": "( \\theta _ { \\mathrm { f a s t } } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 373,
+ 1609,
+ 400,
+ 1609,
+ 400,
+ 1635,
+ 373,
+ 1635
+ ],
+ "score": 0.84,
+ "latex": "X"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 634,
+ 646,
+ 656,
+ 646,
+ 656,
+ 674,
+ 634,
+ 674
+ ],
+ "score": 0.81,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 474,
+ 1204,
+ 491,
+ 1204,
+ 491,
+ 1230,
+ 474,
+ 1230
+ ],
+ "score": 0.8,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 944,
+ 1983,
+ 955,
+ 1983,
+ 955,
+ 2002,
+ 944,
+ 2002
+ ],
+ "score": 0.8,
+ "latex": "t"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 298,
+ 1782,
+ 322,
+ 1782,
+ 322,
+ 1806,
+ 298,
+ 1806
+ ],
+ "score": 0.79,
+ "latex": "C"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1341,
+ 1082,
+ 1372,
+ 1082,
+ 1372,
+ 1112,
+ 1341,
+ 1112
+ ],
+ "score": 0.78,
+ "latex": "\\ S 3"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 572,
+ 232,
+ 600,
+ 232,
+ 600,
+ 259,
+ 572,
+ 259
+ ],
+ "score": 0.78,
+ "latex": "X"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1362,
+ 265,
+ 1383,
+ 265,
+ 1383,
+ 294,
+ 1362,
+ 294
+ ],
+ "score": 0.78,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 553,
+ 1752,
+ 566,
+ 1752,
+ 566,
+ 1776,
+ 553,
+ 1776
+ ],
+ "score": 0.76,
+ "latex": "i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1116,
+ 1609,
+ 1141,
+ 1609,
+ 1141,
+ 1635,
+ 1116,
+ 1635
+ ],
+ "score": 0.75,
+ "latex": "C"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 877,
+ 1644,
+ 896,
+ 1644,
+ 896,
+ 1670,
+ 877,
+ 1670
+ ],
+ "score": 0.75,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1070,
+ 1921,
+ 1089,
+ 1921,
+ 1089,
+ 1944,
+ 1070,
+ 1944
+ ],
+ "score": 0.72,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 696,
+ 582,
+ 713,
+ 582,
+ 713,
+ 608,
+ 696,
+ 608
+ ],
+ "score": 0.71,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 745.0,
+ 393.0,
+ 811.0,
+ 393.0,
+ 811.0,
+ 424.0,
+ 745.0,
+ 424.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 818.0,
+ 379.0,
+ 865.0,
+ 379.0,
+ 865.0,
+ 420.0,
+ 818.0,
+ 420.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 868.0,
+ 379.0,
+ 964.0,
+ 379.0,
+ 964.0,
+ 420.0,
+ 868.0,
+ 420.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 971.0,
+ 385.0,
+ 1007.0,
+ 385.0,
+ 1007.0,
+ 422.0,
+ 971.0,
+ 422.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1015.0,
+ 381.0,
+ 1031.0,
+ 381.0,
+ 1031.0,
+ 414.0,
+ 1015.0,
+ 414.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1035.0,
+ 381.0,
+ 1056.0,
+ 381.0,
+ 1056.0,
+ 418.0,
+ 1035.0,
+ 418.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1059.0,
+ 383.0,
+ 1075.0,
+ 383.0,
+ 1075.0,
+ 417.0,
+ 1059.0,
+ 417.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1206.0,
+ 396.0,
+ 1248.0,
+ 396.0,
+ 1248.0,
+ 417.0,
+ 1206.0,
+ 417.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1264.0,
+ 401.0,
+ 1331.0,
+ 401.0,
+ 1331.0,
+ 428.0,
+ 1264.0,
+ 428.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 818.0,
+ 414.0,
+ 914.0,
+ 414.0,
+ 914.0,
+ 437.0,
+ 818.0,
+ 437.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 930.0,
+ 416.0,
+ 951.0,
+ 416.0,
+ 951.0,
+ 436.0,
+ 930.0,
+ 436.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1016.0,
+ 419.0,
+ 1027.0,
+ 419.0,
+ 1027.0,
+ 432.0,
+ 1016.0,
+ 432.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1036.0,
+ 418.0,
+ 1056.0,
+ 418.0,
+ 1056.0,
+ 434.0,
+ 1036.0,
+ 434.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1059.0,
+ 418.0,
+ 1075.0,
+ 418.0,
+ 1075.0,
+ 434.0,
+ 1059.0,
+ 434.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1206.0,
+ 407.0,
+ 1247.0,
+ 407.0,
+ 1247.0,
+ 427.0,
+ 1206.0,
+ 427.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1147.0,
+ 464.0,
+ 1164.0,
+ 464.0,
+ 1164.0,
+ 483.0,
+ 1147.0,
+ 483.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1079.0,
+ 474.0,
+ 1095.0,
+ 474.0,
+ 1095.0,
+ 491.0,
+ 1079.0,
+ 491.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1143.0,
+ 482.0,
+ 1159.0,
+ 482.0,
+ 1159.0,
+ 501.0,
+ 1143.0,
+ 501.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 825.0,
+ 521.0,
+ 1002.0,
+ 521.0,
+ 1002.0,
+ 586.0,
+ 825.0,
+ 586.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1059.0,
+ 509.0,
+ 1087.0,
+ 509.0,
+ 1087.0,
+ 582.0,
+ 1059.0,
+ 582.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1118.0,
+ 501.0,
+ 1165.0,
+ 501.0,
+ 1165.0,
+ 591.0,
+ 1118.0,
+ 591.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1218.0,
+ 541.0,
+ 1242.0,
+ 541.0,
+ 1242.0,
+ 602.0,
+ 1218.0,
+ 602.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1262.0,
+ 573.0,
+ 1267.0,
+ 573.0,
+ 1267.0,
+ 578.0,
+ 1262.0,
+ 578.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1318.0,
+ 500.0,
+ 1357.0,
+ 500.0,
+ 1357.0,
+ 593.0,
+ 1318.0,
+ 593.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1361.0,
+ 508.0,
+ 1381.0,
+ 508.0,
+ 1381.0,
+ 586.0,
+ 1361.0,
+ 586.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 816.0,
+ 592.0,
+ 988.0,
+ 592.0,
+ 988.0,
+ 625.0,
+ 816.0,
+ 625.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1066.0,
+ 605.0,
+ 1087.0,
+ 605.0,
+ 1087.0,
+ 636.0,
+ 1066.0,
+ 636.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1119.0,
+ 590.0,
+ 1159.0,
+ 590.0,
+ 1159.0,
+ 646.0,
+ 1119.0,
+ 646.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1221.0,
+ 606.0,
+ 1235.0,
+ 606.0,
+ 1235.0,
+ 634.0,
+ 1221.0,
+ 634.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1256.0,
+ 624.0,
+ 1269.0,
+ 624.0,
+ 1269.0,
+ 636.0,
+ 1256.0,
+ 636.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1320.0,
+ 610.0,
+ 1356.0,
+ 610.0,
+ 1356.0,
+ 647.0,
+ 1320.0,
+ 647.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1361.0,
+ 618.0,
+ 1381.0,
+ 618.0,
+ 1381.0,
+ 640.0,
+ 1361.0,
+ 640.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1131.0,
+ 648.0,
+ 1144.0,
+ 648.0,
+ 1144.0,
+ 663.0,
+ 1131.0,
+ 663.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1065.0,
+ 659.0,
+ 1076.0,
+ 659.0,
+ 1076.0,
+ 672.0,
+ 1065.0,
+ 672.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1128.0,
+ 667.0,
+ 1140.0,
+ 667.0,
+ 1140.0,
+ 680.0,
+ 1128.0,
+ 680.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 844.0,
+ 709.0,
+ 1160.0,
+ 709.0,
+ 1160.0,
+ 728.0,
+ 844.0,
+ 728.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1210.0,
+ 708.0,
+ 1255.0,
+ 708.0,
+ 1255.0,
+ 729.0,
+ 1210.0,
+ 729.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 737.0,
+ 763.0,
+ 1045.0,
+ 763.0,
+ 1045.0,
+ 787.0,
+ 737.0,
+ 787.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1113.0,
+ 763.0,
+ 1120.0,
+ 763.0,
+ 1120.0,
+ 787.0,
+ 1113.0,
+ 787.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1186.0,
+ 763.0,
+ 1403.0,
+ 763.0,
+ 1403.0,
+ 787.0,
+ 1186.0,
+ 787.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 735.0,
+ 786.0,
+ 1065.0,
+ 786.0,
+ 1065.0,
+ 813.0,
+ 735.0,
+ 813.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1142.0,
+ 786.0,
+ 1405.0,
+ 786.0,
+ 1405.0,
+ 813.0,
+ 1142.0,
+ 813.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 735.0,
+ 809.0,
+ 1405.0,
+ 809.0,
+ 1405.0,
+ 836.0,
+ 735.0,
+ 836.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 737.0,
+ 834.0,
+ 782.0,
+ 834.0,
+ 782.0,
+ 858.0,
+ 737.0,
+ 858.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 880.0,
+ 834.0,
+ 1403.0,
+ 834.0,
+ 1403.0,
+ 858.0,
+ 880.0,
+ 858.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 735.0,
+ 856.0,
+ 997.0,
+ 856.0,
+ 997.0,
+ 879.0,
+ 735.0,
+ 879.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 330.0,
+ 1972.0,
+ 943.0,
+ 1972.0,
+ 943.0,
+ 2011.0,
+ 330.0,
+ 2011.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 956.0,
+ 1972.0,
+ 1404.0,
+ 1972.0,
+ 1404.0,
+ 2011.0,
+ 956.0,
+ 2011.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 2004.0,
+ 598.0,
+ 2004.0,
+ 598.0,
+ 2034.0,
+ 296.0,
+ 2034.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 989.0,
+ 849.0,
+ 989.0,
+ 849.0,
+ 1025.0,
+ 295.0,
+ 1025.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 72.0,
+ 858.0,
+ 72.0,
+ 858.0,
+ 108.0,
+ 297.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 838.0,
+ 2086.0,
+ 862.0,
+ 2086.0,
+ 862.0,
+ 2118.0,
+ 838.0,
+ 2118.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1049.0,
+ 1406.0,
+ 1049.0,
+ 1406.0,
+ 1086.0,
+ 295.0,
+ 1086.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1080.0,
+ 1340.0,
+ 1080.0,
+ 1340.0,
+ 1117.0,
+ 294.0,
+ 1117.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1373.0,
+ 1080.0,
+ 1406.0,
+ 1080.0,
+ 1406.0,
+ 1117.0,
+ 1373.0,
+ 1117.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1111.0,
+ 977.0,
+ 1111.0,
+ 977.0,
+ 1147.0,
+ 295.0,
+ 1147.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1033.0,
+ 1111.0,
+ 1211.0,
+ 1111.0,
+ 1211.0,
+ 1147.0,
+ 1033.0,
+ 1147.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1258.0,
+ 1111.0,
+ 1407.0,
+ 1111.0,
+ 1407.0,
+ 1147.0,
+ 1258.0,
+ 1147.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1137.0,
+ 456.0,
+ 1137.0,
+ 456.0,
+ 1180.0,
+ 292.0,
+ 1180.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 634.0,
+ 1137.0,
+ 1200.0,
+ 1137.0,
+ 1200.0,
+ 1180.0,
+ 634.0,
+ 1180.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1399.0,
+ 1137.0,
+ 1410.0,
+ 1137.0,
+ 1410.0,
+ 1180.0,
+ 1399.0,
+ 1180.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1171.0,
+ 1407.0,
+ 1171.0,
+ 1407.0,
+ 1208.0,
+ 294.0,
+ 1208.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1201.0,
+ 473.0,
+ 1201.0,
+ 473.0,
+ 1238.0,
+ 294.0,
+ 1238.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 492.0,
+ 1201.0,
+ 1405.0,
+ 1201.0,
+ 1405.0,
+ 1238.0,
+ 492.0,
+ 1238.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1233.0,
+ 1405.0,
+ 1233.0,
+ 1405.0,
+ 1269.0,
+ 294.0,
+ 1269.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1264.0,
+ 1406.0,
+ 1264.0,
+ 1406.0,
+ 1297.0,
+ 295.0,
+ 1297.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1296.0,
+ 765.0,
+ 1296.0,
+ 765.0,
+ 1328.0,
+ 296.0,
+ 1328.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 1342.0,
+ 1405.0,
+ 1342.0,
+ 1405.0,
+ 1375.0,
+ 297.0,
+ 1375.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1371.0,
+ 560.0,
+ 1371.0,
+ 560.0,
+ 1408.0,
+ 294.0,
+ 1408.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 619.0,
+ 1371.0,
+ 1406.0,
+ 1371.0,
+ 1406.0,
+ 1408.0,
+ 619.0,
+ 1408.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1401.0,
+ 494.0,
+ 1401.0,
+ 494.0,
+ 1438.0,
+ 293.0,
+ 1438.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 592.0,
+ 1401.0,
+ 1408.0,
+ 1401.0,
+ 1408.0,
+ 1438.0,
+ 592.0,
+ 1438.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1429.0,
+ 1277.0,
+ 1429.0,
+ 1277.0,
+ 1469.0,
+ 293.0,
+ 1469.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1332.0,
+ 1429.0,
+ 1406.0,
+ 1429.0,
+ 1406.0,
+ 1469.0,
+ 1332.0,
+ 1469.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1458.0,
+ 658.0,
+ 1458.0,
+ 658.0,
+ 1499.0,
+ 292.0,
+ 1499.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 330.0,
+ 721.0,
+ 330.0,
+ 721.0,
+ 362.0,
+ 294.0,
+ 362.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 359.0,
+ 720.0,
+ 359.0,
+ 720.0,
+ 393.0,
+ 296.0,
+ 393.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 388.0,
+ 397.0,
+ 388.0,
+ 397.0,
+ 428.0,
+ 294.0,
+ 428.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 546.0,
+ 426.0,
+ 628.0,
+ 426.0,
+ 628.0,
+ 462.0,
+ 546.0,
+ 462.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 688.0,
+ 426.0,
+ 717.0,
+ 426.0,
+ 717.0,
+ 462.0,
+ 688.0,
+ 462.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 456.0,
+ 719.0,
+ 456.0,
+ 719.0,
+ 492.0,
+ 294.0,
+ 492.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 486.0,
+ 343.0,
+ 486.0,
+ 343.0,
+ 527.0,
+ 293.0,
+ 527.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 370.0,
+ 486.0,
+ 426.0,
+ 486.0,
+ 426.0,
+ 527.0,
+ 370.0,
+ 527.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 709.0,
+ 486.0,
+ 720.0,
+ 486.0,
+ 720.0,
+ 527.0,
+ 709.0,
+ 527.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 516.0,
+ 718.0,
+ 516.0,
+ 718.0,
+ 555.0,
+ 294.0,
+ 555.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 550.0,
+ 716.0,
+ 550.0,
+ 716.0,
+ 582.0,
+ 296.0,
+ 582.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 581.0,
+ 695.0,
+ 581.0,
+ 695.0,
+ 613.0,
+ 292.0,
+ 613.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 714.0,
+ 581.0,
+ 718.0,
+ 581.0,
+ 718.0,
+ 613.0,
+ 714.0,
+ 613.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 613.0,
+ 718.0,
+ 613.0,
+ 718.0,
+ 641.0,
+ 296.0,
+ 641.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 640.0,
+ 633.0,
+ 640.0,
+ 633.0,
+ 673.0,
+ 296.0,
+ 673.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 657.0,
+ 640.0,
+ 719.0,
+ 640.0,
+ 719.0,
+ 673.0,
+ 657.0,
+ 673.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 672.0,
+ 717.0,
+ 672.0,
+ 717.0,
+ 707.0,
+ 295.0,
+ 707.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 703.0,
+ 715.0,
+ 703.0,
+ 715.0,
+ 735.0,
+ 297.0,
+ 735.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 731.0,
+ 721.0,
+ 731.0,
+ 721.0,
+ 765.0,
+ 295.0,
+ 765.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 760.0,
+ 718.0,
+ 760.0,
+ 718.0,
+ 797.0,
+ 294.0,
+ 797.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 795.0,
+ 720.0,
+ 795.0,
+ 720.0,
+ 824.0,
+ 293.0,
+ 824.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 823.0,
+ 717.0,
+ 823.0,
+ 717.0,
+ 857.0,
+ 295.0,
+ 857.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 851.0,
+ 718.0,
+ 851.0,
+ 718.0,
+ 890.0,
+ 293.0,
+ 890.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 882.0,
+ 720.0,
+ 882.0,
+ 720.0,
+ 921.0,
+ 295.0,
+ 921.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 229.0,
+ 571.0,
+ 229.0,
+ 571.0,
+ 265.0,
+ 293.0,
+ 265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 601.0,
+ 229.0,
+ 1406.0,
+ 229.0,
+ 1406.0,
+ 265.0,
+ 601.0,
+ 265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 260.0,
+ 381.0,
+ 260.0,
+ 381.0,
+ 297.0,
+ 293.0,
+ 297.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 500.0,
+ 260.0,
+ 1361.0,
+ 260.0,
+ 1361.0,
+ 297.0,
+ 500.0,
+ 297.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1384.0,
+ 260.0,
+ 1405.0,
+ 260.0,
+ 1405.0,
+ 297.0,
+ 1384.0,
+ 297.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 292.0,
+ 923.0,
+ 292.0,
+ 923.0,
+ 326.0,
+ 295.0,
+ 326.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1848.0,
+ 1405.0,
+ 1848.0,
+ 1405.0,
+ 1886.0,
+ 294.0,
+ 1886.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1882.0,
+ 1407.0,
+ 1882.0,
+ 1407.0,
+ 1916.0,
+ 295.0,
+ 1916.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1911.0,
+ 1069.0,
+ 1911.0,
+ 1069.0,
+ 1949.0,
+ 294.0,
+ 1949.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1090.0,
+ 1911.0,
+ 1409.0,
+ 1911.0,
+ 1409.0,
+ 1949.0,
+ 1090.0,
+ 1949.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1607.0,
+ 372.0,
+ 1607.0,
+ 372.0,
+ 1643.0,
+ 296.0,
+ 1643.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 401.0,
+ 1607.0,
+ 1051.0,
+ 1607.0,
+ 1051.0,
+ 1643.0,
+ 401.0,
+ 1643.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1105.0,
+ 1607.0,
+ 1115.0,
+ 1607.0,
+ 1115.0,
+ 1643.0,
+ 1105.0,
+ 1643.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1142.0,
+ 1607.0,
+ 1405.0,
+ 1607.0,
+ 1405.0,
+ 1643.0,
+ 1142.0,
+ 1643.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1639.0,
+ 876.0,
+ 1639.0,
+ 876.0,
+ 1671.0,
+ 296.0,
+ 1671.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 897.0,
+ 1639.0,
+ 1362.0,
+ 1639.0,
+ 1362.0,
+ 1671.0,
+ 897.0,
+ 1671.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1747.0,
+ 552.0,
+ 1747.0,
+ 552.0,
+ 1784.0,
+ 294.0,
+ 1784.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 567.0,
+ 1747.0,
+ 696.0,
+ 1747.0,
+ 696.0,
+ 1784.0,
+ 567.0,
+ 1784.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 731.0,
+ 1747.0,
+ 891.0,
+ 1747.0,
+ 891.0,
+ 1784.0,
+ 731.0,
+ 1784.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 920.0,
+ 1747.0,
+ 1404.0,
+ 1747.0,
+ 1404.0,
+ 1784.0,
+ 920.0,
+ 1784.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1768.0,
+ 297.0,
+ 1768.0,
+ 297.0,
+ 1823.0,
+ 293.0,
+ 1823.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 323.0,
+ 1768.0,
+ 717.0,
+ 1768.0,
+ 717.0,
+ 1823.0,
+ 323.0,
+ 1823.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 816.0,
+ 1768.0,
+ 830.0,
+ 1768.0,
+ 830.0,
+ 1823.0,
+ 816.0,
+ 1823.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 914.0,
+ 1075.0,
+ 914.0,
+ 1075.0,
+ 949.0,
+ 296.0,
+ 949.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 3,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 750,
+ 1406,
+ 750,
+ 1406,
+ 1029,
+ 297,
+ 1029
+ ],
+ "score": 0.983
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1571,
+ 1404,
+ 1571,
+ 1404,
+ 1787,
+ 298,
+ 1787
+ ],
+ "score": 0.982
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1137,
+ 1405,
+ 1137,
+ 1405,
+ 1321,
+ 298,
+ 1321
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1880,
+ 1402,
+ 1880,
+ 1402,
+ 2035,
+ 299,
+ 2035
+ ],
+ "score": 0.976
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1416,
+ 1404,
+ 1416,
+ 1404,
+ 1539,
+ 298,
+ 1539
+ ],
+ "score": 0.976
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 649,
+ 275,
+ 1047,
+ 275,
+ 1047,
+ 363,
+ 649,
+ 363
+ ],
+ "score": 0.955
+ },
+ {
+ "category_id": 8,
+ "poly": [
+ 520,
+ 539,
+ 1179,
+ 539,
+ 1179,
+ 613,
+ 520,
+ 613
+ ],
+ "score": 0.954
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 296,
+ 463,
+ 1405,
+ 463,
+ 1405,
+ 527,
+ 296,
+ 527
+ ],
+ "score": 0.952
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 300,
+ 376,
+ 1399,
+ 376,
+ 1399,
+ 440,
+ 300,
+ 440
+ ],
+ "score": 0.95
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 300,
+ 1041,
+ 1402,
+ 1041,
+ 1402,
+ 1105,
+ 300,
+ 1105
+ ],
+ "score": 0.949
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 292,
+ 625,
+ 1293,
+ 625,
+ 1293,
+ 659,
+ 292,
+ 659
+ ],
+ "score": 0.933
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 297,
+ 74,
+ 858,
+ 74,
+ 858,
+ 106,
+ 297,
+ 106
+ ],
+ "score": 0.924
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 292,
+ 228,
+ 1288,
+ 228,
+ 1288,
+ 263,
+ 292,
+ 263
+ ],
+ "score": 0.916
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 695,
+ 628,
+ 695,
+ 628,
+ 726,
+ 300,
+ 726
+ ],
+ "score": 0.914
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 1358,
+ 562,
+ 1358,
+ 562,
+ 1390,
+ 300,
+ 1390
+ ],
+ "score": 0.909
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 1824,
+ 641,
+ 1824,
+ 641,
+ 1855,
+ 300,
+ 1855
+ ],
+ "score": 0.902
+ },
+ {
+ "category_id": 9,
+ "poly": [
+ 1366,
+ 308,
+ 1400,
+ 308,
+ 1400,
+ 337,
+ 1366,
+ 337
+ ],
+ "score": 0.881
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 858,
+ 2088,
+ 858,
+ 2112,
+ 841,
+ 2112
+ ],
+ "score": 0.736
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 648,
+ 272,
+ 1049,
+ 272,
+ 1049,
+ 362,
+ 648,
+ 362
+ ],
+ "score": 0.95,
+ "latex": "g _ { i j } = \\frac { \\sum _ { k } ( \\sigma ( \\gamma _ { i j } ) - c _ { i j } ^ { ( k ) } ) \\mathcal { L } _ { C , i } ^ { ( k ) } ( X ) } { \\sum _ { k } \\mathcal { L } _ { C , i } ^ { ( k ) } ( X ) }"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 519,
+ 537,
+ 1181,
+ 537,
+ 1181,
+ 616,
+ 519,
+ 616
+ ],
+ "score": 0.94,
+ "latex": "J _ { \\mathrm { D A G } } = \\sum _ { i \\neq j } \\cosh ( \\sigma ( \\gamma _ { i j } ) \\sigma ( \\gamma _ { j i } ) ) , \\quad \\forall i , j \\in \\{ 0 , \\ldots , M - 1 \\}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1014,
+ 935,
+ 1110,
+ 935,
+ 1110,
+ 967,
+ 1014,
+ 967
+ ],
+ "score": 0.92,
+ "latex": "p a ( i , C )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 644,
+ 1882,
+ 717,
+ 1882,
+ 717,
+ 1916,
+ 644,
+ 1916
+ ],
+ "score": 0.92,
+ "latex": "\\sigma ( \\gamma _ { i j } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 970,
+ 752,
+ 1239,
+ 752,
+ 1239,
+ 788,
+ 970,
+ 788
+ ],
+ "score": 0.92,
+ "latex": "X _ { i } : = f _ { i } ( X _ { p a ( i , C ) } , N _ { i } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 657,
+ 844,
+ 748,
+ 844,
+ 748,
+ 872,
+ 657,
+ 872
+ ],
+ "score": 0.91,
+ "latex": "M \\times N"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 540,
+ 409,
+ 593,
+ 409,
+ 593,
+ 438,
+ 540,
+ 438
+ ],
+ "score": 0.91,
+ "latex": "\\theta _ { \\mathrm { s l o w } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 902,
+ 1725,
+ 955,
+ 1725,
+ 955,
+ 1755,
+ 902,
+ 1755
+ ],
+ "score": 0.9,
+ "latex": "D _ { \\mathrm { i n t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1307,
+ 2003,
+ 1361,
+ 2003,
+ 1361,
+ 2035,
+ 1307,
+ 2035
+ ],
+ "score": 0.9,
+ "latex": "\\theta _ { \\mathrm { s l o w } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1179,
+ 1044,
+ 1258,
+ 1044,
+ 1258,
+ 1070,
+ 1179,
+ 1070
+ ],
+ "score": 0.9,
+ "latex": "N = 2"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 829,
+ 1043,
+ 913,
+ 1043,
+ 913,
+ 1071,
+ 829,
+ 1071
+ ],
+ "score": 0.9,
+ "latex": "M = 3"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 751,
+ 874,
+ 920,
+ 874,
+ 920,
+ 907,
+ 751,
+ 907
+ ],
+ "score": 0.9,
+ "latex": "\\operatorname* { m a x } ( 4 M , 4 N )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 720,
+ 967,
+ 757,
+ 967,
+ 757,
+ 999,
+ 720,
+ 999
+ ],
+ "score": 0.89,
+ "latex": "X _ { j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1347,
+ 1605,
+ 1381,
+ 1605,
+ 1381,
+ 1634,
+ 1347,
+ 1634
+ ],
+ "score": 0.89,
+ "latex": "X _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 896,
+ 996,
+ 1275,
+ 996,
+ 1275,
+ 1030,
+ 896,
+ 1030
+ ],
+ "score": 0.89,
+ "latex": "\\theta = \\left\\{ { \\cal W } 0 _ { i h j n } , { \\tt B } 0 _ { i h } , { \\tt W } 1 _ { i n h } , { \\tt B } 1 _ { i n } \\right\\}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 979,
+ 1418,
+ 1013,
+ 1418,
+ 1013,
+ 1447,
+ 979,
+ 1447
+ ],
+ "score": 0.88,
+ "latex": "X _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 891,
+ 1574,
+ 926,
+ 1574,
+ 926,
+ 1603,
+ 891,
+ 1603
+ ],
+ "score": 0.88,
+ "latex": "X _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 897,
+ 966,
+ 931,
+ 966,
+ 931,
+ 996,
+ 897,
+ 996
+ ],
+ "score": 0.87,
+ "latex": "X _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1196,
+ 1418,
+ 1230,
+ 1418,
+ 1230,
+ 1447,
+ 1196,
+ 1447
+ ],
+ "score": 0.87,
+ "latex": "X _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 501,
+ 936,
+ 527,
+ 936,
+ 527,
+ 967,
+ 501,
+ 967
+ ],
+ "score": 0.87,
+ "latex": "f _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 297,
+ 1000,
+ 331,
+ 1000,
+ 331,
+ 1030,
+ 297,
+ 1030
+ ],
+ "score": 0.86,
+ "latex": "c _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 661,
+ 906,
+ 689,
+ 906,
+ 689,
+ 932,
+ 661,
+ 932
+ ],
+ "score": 0.84,
+ "latex": "N"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1018,
+ 379,
+ 1037,
+ 379,
+ 1037,
+ 405,
+ 1018,
+ 405
+ ],
+ "score": 0.83,
+ "latex": "k"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1003,
+ 1235,
+ 1024,
+ 1235,
+ 1024,
+ 1261,
+ 1003,
+ 1261
+ ],
+ "score": 0.82,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 893,
+ 845,
+ 926,
+ 845,
+ 926,
+ 871,
+ 893,
+ 871
+ ],
+ "score": 0.82,
+ "latex": "M"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1355,
+ 1884,
+ 1371,
+ 1884,
+ 1371,
+ 1914,
+ 1355,
+ 1914
+ ],
+ "score": 0.82,
+ "latex": "j"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1173,
+ 379,
+ 1197,
+ 379,
+ 1197,
+ 405,
+ 1173,
+ 405
+ ],
+ "score": 0.82,
+ "latex": "C"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 831,
+ 1264,
+ 852,
+ 1264,
+ 852,
+ 1292,
+ 831,
+ 1292
+ ],
+ "score": 0.82,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1374,
+ 1578,
+ 1400,
+ 1578,
+ 1400,
+ 1604,
+ 1374,
+ 1604
+ ],
+ "score": 0.81,
+ "latex": "\\gamma _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 927,
+ 409,
+ 944,
+ 409,
+ 944,
+ 435,
+ 927,
+ 435
+ ],
+ "score": 0.79,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1209,
+ 845,
+ 1237,
+ 845,
+ 1237,
+ 871,
+ 1209,
+ 871
+ ],
+ "score": 0.79,
+ "latex": "N"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1376,
+ 936,
+ 1401,
+ 936,
+ 1401,
+ 962,
+ 1376,
+ 962
+ ],
+ "score": 0.79,
+ "latex": "C"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 471,
+ 784,
+ 503,
+ 784,
+ 503,
+ 810,
+ 471,
+ 810
+ ],
+ "score": 0.79,
+ "latex": "M"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 671,
+ 754,
+ 704,
+ 754,
+ 704,
+ 780,
+ 671,
+ 780
+ ],
+ "score": 0.78,
+ "latex": "M"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 761,
+ 1261,
+ 779,
+ 1261,
+ 779,
+ 1287,
+ 761,
+ 1287
+ ],
+ "score": 0.78,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 620,
+ 1231,
+ 638,
+ 1231,
+ 638,
+ 1257,
+ 620,
+ 1257
+ ],
+ "score": 0.77,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1235,
+ 1884,
+ 1249,
+ 1884,
+ 1249,
+ 1909,
+ 1235,
+ 1909
+ ],
+ "score": 0.72,
+ "latex": "i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 619,
+ 1606,
+ 632,
+ 1606,
+ 632,
+ 1631,
+ 619,
+ 1631
+ ],
+ "score": 0.66,
+ "latex": "i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 419,
+ 373,
+ 451,
+ 373,
+ 451,
+ 404,
+ 419,
+ 404
+ ],
+ "score": 0.53,
+ "latex": "( k )"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 72.0,
+ 858.0,
+ 72.0,
+ 858.0,
+ 109.0,
+ 297.0,
+ 109.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 692.0,
+ 631.0,
+ 692.0,
+ 631.0,
+ 730.0,
+ 294.0,
+ 730.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1356.0,
+ 564.0,
+ 1356.0,
+ 564.0,
+ 1393.0,
+ 294.0,
+ 1393.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1823.0,
+ 645.0,
+ 1823.0,
+ 645.0,
+ 1859.0,
+ 295.0,
+ 1859.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2085.0,
+ 861.0,
+ 2085.0,
+ 861.0,
+ 2120.0,
+ 839.0,
+ 2120.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 750.0,
+ 670.0,
+ 750.0,
+ 670.0,
+ 791.0,
+ 292.0,
+ 791.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 705.0,
+ 750.0,
+ 969.0,
+ 750.0,
+ 969.0,
+ 791.0,
+ 705.0,
+ 791.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1240.0,
+ 750.0,
+ 1407.0,
+ 750.0,
+ 1407.0,
+ 791.0,
+ 1240.0,
+ 791.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 783.0,
+ 470.0,
+ 783.0,
+ 470.0,
+ 816.0,
+ 296.0,
+ 816.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 504.0,
+ 783.0,
+ 1404.0,
+ 783.0,
+ 1404.0,
+ 816.0,
+ 504.0,
+ 816.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 812.0,
+ 1409.0,
+ 812.0,
+ 1409.0,
+ 849.0,
+ 294.0,
+ 849.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 842.0,
+ 656.0,
+ 842.0,
+ 656.0,
+ 879.0,
+ 294.0,
+ 879.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 749.0,
+ 842.0,
+ 892.0,
+ 842.0,
+ 892.0,
+ 879.0,
+ 749.0,
+ 879.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 927.0,
+ 842.0,
+ 1208.0,
+ 842.0,
+ 1208.0,
+ 879.0,
+ 927.0,
+ 879.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1238.0,
+ 842.0,
+ 1406.0,
+ 842.0,
+ 1406.0,
+ 879.0,
+ 1238.0,
+ 879.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 872.0,
+ 750.0,
+ 872.0,
+ 750.0,
+ 909.0,
+ 294.0,
+ 909.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 921.0,
+ 872.0,
+ 1407.0,
+ 872.0,
+ 1407.0,
+ 909.0,
+ 921.0,
+ 909.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 903.0,
+ 660.0,
+ 903.0,
+ 660.0,
+ 939.0,
+ 294.0,
+ 939.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 690.0,
+ 903.0,
+ 1406.0,
+ 903.0,
+ 1406.0,
+ 939.0,
+ 690.0,
+ 939.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 933.0,
+ 500.0,
+ 933.0,
+ 500.0,
+ 970.0,
+ 292.0,
+ 970.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 528.0,
+ 933.0,
+ 1013.0,
+ 933.0,
+ 1013.0,
+ 970.0,
+ 528.0,
+ 970.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1111.0,
+ 933.0,
+ 1375.0,
+ 933.0,
+ 1375.0,
+ 970.0,
+ 1111.0,
+ 970.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 966.0,
+ 719.0,
+ 966.0,
+ 719.0,
+ 999.0,
+ 296.0,
+ 999.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 758.0,
+ 966.0,
+ 896.0,
+ 966.0,
+ 896.0,
+ 999.0,
+ 758.0,
+ 999.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 932.0,
+ 966.0,
+ 1404.0,
+ 966.0,
+ 1404.0,
+ 999.0,
+ 932.0,
+ 999.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 332.0,
+ 993.0,
+ 895.0,
+ 993.0,
+ 895.0,
+ 1034.0,
+ 332.0,
+ 1034.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1276.0,
+ 993.0,
+ 1285.0,
+ 993.0,
+ 1285.0,
+ 1034.0,
+ 1276.0,
+ 1034.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1568.0,
+ 890.0,
+ 1568.0,
+ 890.0,
+ 1610.0,
+ 292.0,
+ 1610.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 927.0,
+ 1568.0,
+ 1373.0,
+ 1568.0,
+ 1373.0,
+ 1610.0,
+ 927.0,
+ 1610.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1401.0,
+ 1568.0,
+ 1405.0,
+ 1568.0,
+ 1405.0,
+ 1610.0,
+ 1401.0,
+ 1610.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1605.0,
+ 618.0,
+ 1605.0,
+ 618.0,
+ 1636.0,
+ 296.0,
+ 1636.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 633.0,
+ 1605.0,
+ 1346.0,
+ 1605.0,
+ 1346.0,
+ 1636.0,
+ 633.0,
+ 1636.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1382.0,
+ 1605.0,
+ 1405.0,
+ 1605.0,
+ 1405.0,
+ 1636.0,
+ 1382.0,
+ 1636.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1633.0,
+ 1405.0,
+ 1633.0,
+ 1405.0,
+ 1668.0,
+ 294.0,
+ 1668.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1664.0,
+ 1406.0,
+ 1664.0,
+ 1406.0,
+ 1699.0,
+ 294.0,
+ 1699.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1695.0,
+ 1405.0,
+ 1695.0,
+ 1405.0,
+ 1726.0,
+ 293.0,
+ 1726.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1723.0,
+ 901.0,
+ 1723.0,
+ 901.0,
+ 1759.0,
+ 293.0,
+ 1759.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 956.0,
+ 1723.0,
+ 1406.0,
+ 1723.0,
+ 1406.0,
+ 1759.0,
+ 956.0,
+ 1759.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1755.0,
+ 742.0,
+ 1755.0,
+ 742.0,
+ 1790.0,
+ 295.0,
+ 1790.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1138.0,
+ 1407.0,
+ 1138.0,
+ 1407.0,
+ 1173.0,
+ 296.0,
+ 1173.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1169.0,
+ 1403.0,
+ 1169.0,
+ 1403.0,
+ 1201.0,
+ 295.0,
+ 1201.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1197.0,
+ 1408.0,
+ 1197.0,
+ 1408.0,
+ 1234.0,
+ 292.0,
+ 1234.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1228.0,
+ 619.0,
+ 1228.0,
+ 619.0,
+ 1265.0,
+ 293.0,
+ 1265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 639.0,
+ 1228.0,
+ 1002.0,
+ 1228.0,
+ 1002.0,
+ 1265.0,
+ 639.0,
+ 1265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1025.0,
+ 1228.0,
+ 1407.0,
+ 1228.0,
+ 1407.0,
+ 1265.0,
+ 1025.0,
+ 1265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1258.0,
+ 760.0,
+ 1258.0,
+ 760.0,
+ 1295.0,
+ 292.0,
+ 1295.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 780.0,
+ 1258.0,
+ 830.0,
+ 1258.0,
+ 830.0,
+ 1295.0,
+ 780.0,
+ 1295.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 853.0,
+ 1258.0,
+ 1407.0,
+ 1258.0,
+ 1407.0,
+ 1295.0,
+ 853.0,
+ 1295.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1292.0,
+ 684.0,
+ 1292.0,
+ 684.0,
+ 1324.0,
+ 296.0,
+ 1324.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1880.0,
+ 643.0,
+ 1880.0,
+ 643.0,
+ 1918.0,
+ 294.0,
+ 1918.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 718.0,
+ 1880.0,
+ 1234.0,
+ 1880.0,
+ 1234.0,
+ 1918.0,
+ 718.0,
+ 1918.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1250.0,
+ 1880.0,
+ 1354.0,
+ 1880.0,
+ 1354.0,
+ 1918.0,
+ 1250.0,
+ 1918.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1372.0,
+ 1880.0,
+ 1407.0,
+ 1880.0,
+ 1407.0,
+ 1918.0,
+ 1372.0,
+ 1918.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1910.0,
+ 1406.0,
+ 1910.0,
+ 1406.0,
+ 1947.0,
+ 294.0,
+ 1947.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1942.0,
+ 1406.0,
+ 1942.0,
+ 1406.0,
+ 1979.0,
+ 294.0,
+ 1979.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1971.0,
+ 1406.0,
+ 1971.0,
+ 1406.0,
+ 2008.0,
+ 294.0,
+ 2008.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1999.0,
+ 1306.0,
+ 1999.0,
+ 1306.0,
+ 2041.0,
+ 293.0,
+ 2041.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1362.0,
+ 1999.0,
+ 1407.0,
+ 1999.0,
+ 1407.0,
+ 2041.0,
+ 1362.0,
+ 2041.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1415.0,
+ 978.0,
+ 1415.0,
+ 978.0,
+ 1450.0,
+ 293.0,
+ 1450.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1014.0,
+ 1415.0,
+ 1195.0,
+ 1415.0,
+ 1195.0,
+ 1450.0,
+ 1014.0,
+ 1450.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1231.0,
+ 1415.0,
+ 1402.0,
+ 1415.0,
+ 1402.0,
+ 1450.0,
+ 1231.0,
+ 1450.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1446.0,
+ 1402.0,
+ 1446.0,
+ 1402.0,
+ 1479.0,
+ 294.0,
+ 1479.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1479.0,
+ 1406.0,
+ 1479.0,
+ 1406.0,
+ 1511.0,
+ 292.0,
+ 1511.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1508.0,
+ 474.0,
+ 1508.0,
+ 474.0,
+ 1541.0,
+ 293.0,
+ 1541.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 463.0,
+ 1403.0,
+ 463.0,
+ 1403.0,
+ 499.0,
+ 296.0,
+ 499.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 495.0,
+ 1177.0,
+ 495.0,
+ 1177.0,
+ 531.0,
+ 294.0,
+ 531.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 368.0,
+ 418.0,
+ 368.0,
+ 418.0,
+ 414.0,
+ 292.0,
+ 414.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 452.0,
+ 368.0,
+ 1017.0,
+ 368.0,
+ 1017.0,
+ 414.0,
+ 452.0,
+ 414.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1038.0,
+ 368.0,
+ 1172.0,
+ 368.0,
+ 1172.0,
+ 414.0,
+ 1038.0,
+ 414.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1198.0,
+ 368.0,
+ 1408.0,
+ 368.0,
+ 1408.0,
+ 414.0,
+ 1198.0,
+ 414.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 405.0,
+ 539.0,
+ 405.0,
+ 539.0,
+ 444.0,
+ 295.0,
+ 444.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 594.0,
+ 405.0,
+ 926.0,
+ 405.0,
+ 926.0,
+ 444.0,
+ 594.0,
+ 444.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 945.0,
+ 405.0,
+ 1264.0,
+ 405.0,
+ 1264.0,
+ 444.0,
+ 945.0,
+ 444.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1041.0,
+ 828.0,
+ 1041.0,
+ 828.0,
+ 1077.0,
+ 296.0,
+ 1077.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 914.0,
+ 1041.0,
+ 1178.0,
+ 1041.0,
+ 1178.0,
+ 1077.0,
+ 914.0,
+ 1077.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1259.0,
+ 1041.0,
+ 1406.0,
+ 1041.0,
+ 1406.0,
+ 1077.0,
+ 1259.0,
+ 1077.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1073.0,
+ 512.0,
+ 1073.0,
+ 512.0,
+ 1106.0,
+ 296.0,
+ 1106.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 622.0,
+ 1297.0,
+ 622.0,
+ 1297.0,
+ 663.0,
+ 294.0,
+ 663.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 224.0,
+ 1294.0,
+ 224.0,
+ 1294.0,
+ 271.0,
+ 292.0,
+ 271.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 4,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 857,
+ 1406,
+ 857,
+ 1406,
+ 1315,
+ 298,
+ 1315
+ ],
+ "score": 0.984
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1758,
+ 1404,
+ 1758,
+ 1404,
+ 2034,
+ 298,
+ 2034
+ ],
+ "score": 0.982
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1498,
+ 1403,
+ 1498,
+ 1403,
+ 1744,
+ 298,
+ 1744
+ ],
+ "score": 0.981
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 537,
+ 1404,
+ 537,
+ 1404,
+ 783,
+ 298,
+ 783
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 338,
+ 1404,
+ 338,
+ 1404,
+ 522,
+ 298,
+ 522
+ ],
+ "score": 0.978
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1330,
+ 1404,
+ 1330,
+ 1404,
+ 1483,
+ 298,
+ 1483
+ ],
+ "score": 0.977
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 301,
+ 229,
+ 1401,
+ 229,
+ 1401,
+ 323,
+ 301,
+ 323
+ ],
+ "score": 0.965
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 808,
+ 588,
+ 808,
+ 588,
+ 843,
+ 300,
+ 843
+ ],
+ "score": 0.905
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 76,
+ 857,
+ 76,
+ 857,
+ 104,
+ 298,
+ 104
+ ],
+ "score": 0.887
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 840,
+ 2089,
+ 858,
+ 2089,
+ 858,
+ 2112,
+ 840,
+ 2112
+ ],
+ "score": 0.779
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 754,
+ 598,
+ 880,
+ 598,
+ 880,
+ 632,
+ 754,
+ 632
+ ],
+ "score": 0.93,
+ "latex": "\\sigma ( \\gamma _ { i j } ) \\approx 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 394,
+ 660,
+ 468,
+ 660,
+ 468,
+ 693,
+ 394,
+ 693
+ ],
+ "score": 0.92,
+ "latex": "\\sigma ( \\gamma _ { i j } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 710,
+ 721,
+ 763,
+ 721,
+ 763,
+ 751,
+ 710,
+ 751
+ ],
+ "score": 0.91,
+ "latex": "\\theta _ { \\mathrm { s l o w } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1179,
+ 629,
+ 1216,
+ 629,
+ 1216,
+ 663,
+ 1179,
+ 663
+ ],
+ "score": 0.89,
+ "latex": "X _ { j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 362,
+ 630,
+ 399,
+ 630,
+ 399,
+ 663,
+ 362,
+ 663
+ ],
+ "score": 0.88,
+ "latex": "X _ { j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1233,
+ 343,
+ 1270,
+ 343,
+ 1270,
+ 373,
+ 1233,
+ 373
+ ],
+ "score": 0.87,
+ "latex": "\\gamma _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 953,
+ 345,
+ 972,
+ 345,
+ 972,
+ 371,
+ 953,
+ 371
+ ],
+ "score": 0.8,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 489,
+ 757,
+ 508,
+ 757,
+ 508,
+ 783,
+ 489,
+ 783
+ ],
+ "score": 0.79,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1378,
+ 436,
+ 1397,
+ 436,
+ 1397,
+ 462,
+ 1378,
+ 462
+ ],
+ "score": 0.77,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 651,
+ 631,
+ 664,
+ 631,
+ 664,
+ 656,
+ 651,
+ 656
+ ],
+ "score": 0.76,
+ "latex": "i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 404,
+ 462,
+ 422,
+ 462,
+ 422,
+ 488,
+ 404,
+ 488
+ ],
+ "score": 0.73,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1056,
+ 600,
+ 1069,
+ 600,
+ 1069,
+ 626,
+ 1056,
+ 626
+ ],
+ "score": 0.71,
+ "latex": "i"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 803.0,
+ 593.0,
+ 803.0,
+ 593.0,
+ 849.0,
+ 292.0,
+ 849.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 298.0,
+ 73.0,
+ 858.0,
+ 73.0,
+ 858.0,
+ 108.0,
+ 298.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 840.0,
+ 2087.0,
+ 861.0,
+ 2087.0,
+ 861.0,
+ 2118.0,
+ 840.0,
+ 2118.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 854.0,
+ 1407.0,
+ 854.0,
+ 1407.0,
+ 894.0,
+ 292.0,
+ 894.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 889.0,
+ 1407.0,
+ 889.0,
+ 1407.0,
+ 920.0,
+ 295.0,
+ 920.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 920.0,
+ 1406.0,
+ 920.0,
+ 1406.0,
+ 951.0,
+ 296.0,
+ 951.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 949.0,
+ 1407.0,
+ 949.0,
+ 1407.0,
+ 983.0,
+ 295.0,
+ 983.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 979.0,
+ 1408.0,
+ 979.0,
+ 1408.0,
+ 1013.0,
+ 293.0,
+ 1013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1008.0,
+ 1408.0,
+ 1008.0,
+ 1408.0,
+ 1045.0,
+ 293.0,
+ 1045.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1039.0,
+ 1406.0,
+ 1039.0,
+ 1406.0,
+ 1074.0,
+ 295.0,
+ 1074.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1067.0,
+ 1406.0,
+ 1067.0,
+ 1406.0,
+ 1108.0,
+ 292.0,
+ 1108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1100.0,
+ 1407.0,
+ 1100.0,
+ 1407.0,
+ 1137.0,
+ 292.0,
+ 1137.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1132.0,
+ 1408.0,
+ 1132.0,
+ 1408.0,
+ 1166.0,
+ 293.0,
+ 1166.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1163.0,
+ 1406.0,
+ 1163.0,
+ 1406.0,
+ 1194.0,
+ 296.0,
+ 1194.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1193.0,
+ 1406.0,
+ 1193.0,
+ 1406.0,
+ 1227.0,
+ 293.0,
+ 1227.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1223.0,
+ 1404.0,
+ 1223.0,
+ 1404.0,
+ 1258.0,
+ 295.0,
+ 1258.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1252.0,
+ 1406.0,
+ 1252.0,
+ 1406.0,
+ 1288.0,
+ 293.0,
+ 1288.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1284.0,
+ 1101.0,
+ 1284.0,
+ 1101.0,
+ 1319.0,
+ 295.0,
+ 1319.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1759.0,
+ 1406.0,
+ 1759.0,
+ 1406.0,
+ 1795.0,
+ 296.0,
+ 1795.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1789.0,
+ 1406.0,
+ 1789.0,
+ 1406.0,
+ 1826.0,
+ 294.0,
+ 1826.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1822.0,
+ 1405.0,
+ 1822.0,
+ 1405.0,
+ 1854.0,
+ 296.0,
+ 1854.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1848.0,
+ 1406.0,
+ 1848.0,
+ 1406.0,
+ 1886.0,
+ 293.0,
+ 1886.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1882.0,
+ 1406.0,
+ 1882.0,
+ 1406.0,
+ 1918.0,
+ 294.0,
+ 1918.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1912.0,
+ 1405.0,
+ 1912.0,
+ 1405.0,
+ 1947.0,
+ 293.0,
+ 1947.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1944.0,
+ 1405.0,
+ 1944.0,
+ 1405.0,
+ 1977.0,
+ 294.0,
+ 1977.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1972.0,
+ 1406.0,
+ 1972.0,
+ 1406.0,
+ 2008.0,
+ 294.0,
+ 2008.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 2003.0,
+ 1408.0,
+ 2003.0,
+ 1408.0,
+ 2040.0,
+ 294.0,
+ 2040.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1500.0,
+ 1404.0,
+ 1500.0,
+ 1404.0,
+ 1533.0,
+ 294.0,
+ 1533.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1530.0,
+ 1405.0,
+ 1530.0,
+ 1405.0,
+ 1563.0,
+ 292.0,
+ 1563.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1561.0,
+ 1405.0,
+ 1561.0,
+ 1405.0,
+ 1595.0,
+ 293.0,
+ 1595.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1590.0,
+ 1407.0,
+ 1590.0,
+ 1407.0,
+ 1626.0,
+ 292.0,
+ 1626.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1621.0,
+ 1405.0,
+ 1621.0,
+ 1405.0,
+ 1655.0,
+ 294.0,
+ 1655.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1652.0,
+ 1405.0,
+ 1652.0,
+ 1405.0,
+ 1685.0,
+ 293.0,
+ 1685.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1682.0,
+ 1407.0,
+ 1682.0,
+ 1407.0,
+ 1716.0,
+ 294.0,
+ 1716.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1712.0,
+ 1161.0,
+ 1712.0,
+ 1161.0,
+ 1746.0,
+ 296.0,
+ 1746.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 537.0,
+ 1405.0,
+ 537.0,
+ 1405.0,
+ 571.0,
+ 296.0,
+ 571.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 569.0,
+ 1404.0,
+ 569.0,
+ 1404.0,
+ 602.0,
+ 294.0,
+ 602.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 598.0,
+ 753.0,
+ 598.0,
+ 753.0,
+ 635.0,
+ 293.0,
+ 635.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 881.0,
+ 598.0,
+ 1055.0,
+ 598.0,
+ 1055.0,
+ 635.0,
+ 881.0,
+ 635.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1070.0,
+ 598.0,
+ 1406.0,
+ 598.0,
+ 1406.0,
+ 635.0,
+ 1070.0,
+ 635.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 630.0,
+ 361.0,
+ 630.0,
+ 361.0,
+ 664.0,
+ 294.0,
+ 664.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 400.0,
+ 630.0,
+ 650.0,
+ 630.0,
+ 650.0,
+ 664.0,
+ 400.0,
+ 664.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 665.0,
+ 630.0,
+ 1178.0,
+ 630.0,
+ 1178.0,
+ 664.0,
+ 665.0,
+ 664.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1217.0,
+ 630.0,
+ 1405.0,
+ 630.0,
+ 1405.0,
+ 664.0,
+ 1217.0,
+ 664.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 659.0,
+ 393.0,
+ 659.0,
+ 393.0,
+ 694.0,
+ 292.0,
+ 694.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 469.0,
+ 659.0,
+ 1407.0,
+ 659.0,
+ 1407.0,
+ 694.0,
+ 469.0,
+ 694.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 688.0,
+ 1405.0,
+ 688.0,
+ 1405.0,
+ 724.0,
+ 293.0,
+ 724.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 719.0,
+ 709.0,
+ 719.0,
+ 709.0,
+ 755.0,
+ 293.0,
+ 755.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 764.0,
+ 719.0,
+ 1404.0,
+ 719.0,
+ 1404.0,
+ 755.0,
+ 764.0,
+ 755.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 750.0,
+ 488.0,
+ 750.0,
+ 488.0,
+ 789.0,
+ 294.0,
+ 789.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 509.0,
+ 750.0,
+ 520.0,
+ 750.0,
+ 520.0,
+ 789.0,
+ 509.0,
+ 789.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 337.0,
+ 952.0,
+ 337.0,
+ 952.0,
+ 373.0,
+ 296.0,
+ 373.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 973.0,
+ 337.0,
+ 1232.0,
+ 337.0,
+ 1232.0,
+ 373.0,
+ 973.0,
+ 373.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1271.0,
+ 337.0,
+ 1406.0,
+ 337.0,
+ 1406.0,
+ 373.0,
+ 1271.0,
+ 373.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 368.0,
+ 1406.0,
+ 368.0,
+ 1406.0,
+ 403.0,
+ 293.0,
+ 403.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 401.0,
+ 1405.0,
+ 401.0,
+ 1405.0,
+ 433.0,
+ 296.0,
+ 433.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 429.0,
+ 1377.0,
+ 429.0,
+ 1377.0,
+ 468.0,
+ 293.0,
+ 468.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1398.0,
+ 429.0,
+ 1409.0,
+ 429.0,
+ 1409.0,
+ 468.0,
+ 1398.0,
+ 468.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 460.0,
+ 403.0,
+ 460.0,
+ 403.0,
+ 492.0,
+ 296.0,
+ 492.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 423.0,
+ 460.0,
+ 1402.0,
+ 460.0,
+ 1402.0,
+ 492.0,
+ 423.0,
+ 492.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 491.0,
+ 982.0,
+ 491.0,
+ 982.0,
+ 524.0,
+ 293.0,
+ 524.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1329.0,
+ 1408.0,
+ 1329.0,
+ 1408.0,
+ 1364.0,
+ 293.0,
+ 1364.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1360.0,
+ 1405.0,
+ 1360.0,
+ 1405.0,
+ 1393.0,
+ 296.0,
+ 1393.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1388.0,
+ 1406.0,
+ 1388.0,
+ 1406.0,
+ 1428.0,
+ 293.0,
+ 1428.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1421.0,
+ 1404.0,
+ 1421.0,
+ 1404.0,
+ 1457.0,
+ 293.0,
+ 1457.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1451.0,
+ 1406.0,
+ 1451.0,
+ 1406.0,
+ 1487.0,
+ 295.0,
+ 1487.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 230.0,
+ 1404.0,
+ 230.0,
+ 1404.0,
+ 264.0,
+ 296.0,
+ 264.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 261.0,
+ 1405.0,
+ 261.0,
+ 1405.0,
+ 295.0,
+ 296.0,
+ 295.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 291.0,
+ 490.0,
+ 291.0,
+ 490.0,
+ 326.0,
+ 295.0,
+ 326.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 5,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 278,
+ 1405,
+ 278,
+ 1405,
+ 554,
+ 297,
+ 554
+ ],
+ "score": 0.982
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1130,
+ 1405,
+ 1130,
+ 1405,
+ 1345,
+ 297,
+ 1345
+ ],
+ "score": 0.981
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 823,
+ 1405,
+ 823,
+ 1405,
+ 1098,
+ 297,
+ 1098
+ ],
+ "score": 0.981
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1820,
+ 989,
+ 1820,
+ 989,
+ 2033,
+ 298,
+ 2033
+ ],
+ "score": 0.979
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 1437,
+ 993,
+ 1437,
+ 993,
+ 1651,
+ 299,
+ 1651
+ ],
+ "score": 0.974
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1653,
+ 1405,
+ 1653,
+ 1405,
+ 1803,
+ 298,
+ 1803
+ ],
+ "score": 0.971
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 1017,
+ 1827,
+ 1395,
+ 1827,
+ 1395,
+ 1952,
+ 1017,
+ 1952
+ ],
+ "score": 0.965
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 449,
+ 626,
+ 1242,
+ 626,
+ 1242,
+ 743,
+ 449,
+ 743
+ ],
+ "score": 0.964
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 1015,
+ 1445,
+ 1398,
+ 1445,
+ 1398,
+ 1571,
+ 1015,
+ 1571
+ ],
+ "score": 0.96
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 1012,
+ 1595,
+ 1402,
+ 1595,
+ 1402,
+ 1641,
+ 1012,
+ 1641
+ ],
+ "score": 0.928
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 1013,
+ 1976,
+ 1402,
+ 1976,
+ 1402,
+ 2021,
+ 1013,
+ 2021
+ ],
+ "score": 0.927
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 296,
+ 764,
+ 1401,
+ 764,
+ 1401,
+ 815,
+ 296,
+ 815
+ ],
+ "score": 0.925
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 589,
+ 636,
+ 589,
+ 636,
+ 618,
+ 299,
+ 618
+ ],
+ "score": 0.905
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 75,
+ 858,
+ 75,
+ 858,
+ 104,
+ 298,
+ 104
+ ],
+ "score": 0.882
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 302,
+ 1380,
+ 805,
+ 1380,
+ 805,
+ 1411,
+ 302,
+ 1411
+ ],
+ "score": 0.866
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 228,
+ 878,
+ 228,
+ 878,
+ 261,
+ 299,
+ 261
+ ],
+ "score": 0.859
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 841,
+ 2088,
+ 857,
+ 2088,
+ 857,
+ 2111,
+ 841,
+ 2111
+ ],
+ "score": 0.746
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 297,
+ 1942,
+ 372,
+ 1942,
+ 372,
+ 1976,
+ 297,
+ 1976
+ ],
+ "score": 0.93,
+ "latex": "\\sigma ( \\gamma _ { i j } )"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 703,
+ 854,
+ 787,
+ 854,
+ 787,
+ 882,
+ 703,
+ 882
+ ],
+ "score": 0.91,
+ "latex": "M = 3"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 612,
+ 1192,
+ 702,
+ 1192,
+ 702,
+ 1220,
+ 612,
+ 1220
+ ],
+ "score": 0.9,
+ "latex": "M = 3"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 337,
+ 976,
+ 421,
+ 976,
+ 421,
+ 1005,
+ 337,
+ 1005
+ ],
+ "score": 0.88,
+ "latex": "M > 3"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 421,
+ 1945,
+ 456,
+ 1945,
+ 456,
+ 1976,
+ 421,
+ 1976
+ ],
+ "score": 0.88,
+ "latex": "c _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1316,
+ 374,
+ 1351,
+ 374,
+ 1351,
+ 403,
+ 1316,
+ 403
+ ],
+ "score": 0.87,
+ "latex": "c _ { i j }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 475,
+ 1007,
+ 567,
+ 1007,
+ 567,
+ 1035,
+ 475,
+ 1035
+ ],
+ "score": 0.85,
+ "latex": "( \\mathbb { M } = 3 - 8"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 479,
+ 1590,
+ 560,
+ 1590,
+ 560,
+ 1619,
+ 479,
+ 1619
+ ],
+ "score": 0.78,
+ "latex": "M = 5"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1115.0,
+ 1848.0,
+ 1122.0,
+ 1848.0,
+ 1122.0,
+ 1856.0,
+ 1115.0,
+ 1856.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1129.0,
+ 1848.0,
+ 1138.0,
+ 1848.0,
+ 1138.0,
+ 1855.0,
+ 1129.0,
+ 1855.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1150.0,
+ 1847.0,
+ 1160.0,
+ 1847.0,
+ 1160.0,
+ 1856.0,
+ 1150.0,
+ 1856.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1181.0,
+ 1848.0,
+ 1189.0,
+ 1848.0,
+ 1189.0,
+ 1855.0,
+ 1181.0,
+ 1855.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1278.0,
+ 1847.0,
+ 1288.0,
+ 1847.0,
+ 1288.0,
+ 1855.0,
+ 1278.0,
+ 1855.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1130.0,
+ 1863.0,
+ 1137.0,
+ 1863.0,
+ 1137.0,
+ 1871.0,
+ 1130.0,
+ 1871.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1151.0,
+ 1863.0,
+ 1159.0,
+ 1863.0,
+ 1159.0,
+ 1871.0,
+ 1151.0,
+ 1871.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1166.0,
+ 1863.0,
+ 1173.0,
+ 1863.0,
+ 1173.0,
+ 1871.0,
+ 1166.0,
+ 1871.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1198.0,
+ 1864.0,
+ 1203.0,
+ 1864.0,
+ 1203.0,
+ 1869.0,
+ 1198.0,
+ 1869.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1242.0,
+ 1863.0,
+ 1250.0,
+ 1863.0,
+ 1250.0,
+ 1870.0,
+ 1242.0,
+ 1870.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1256.0,
+ 1863.0,
+ 1265.0,
+ 1863.0,
+ 1265.0,
+ 1871.0,
+ 1256.0,
+ 1871.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1324.0,
+ 1863.0,
+ 1331.0,
+ 1863.0,
+ 1331.0,
+ 1870.0,
+ 1324.0,
+ 1870.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1151.0,
+ 1909.0,
+ 1159.0,
+ 1909.0,
+ 1159.0,
+ 1916.0,
+ 1151.0,
+ 1916.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1292.0,
+ 1909.0,
+ 1302.0,
+ 1909.0,
+ 1302.0,
+ 1918.0,
+ 1292.0,
+ 1918.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1182.0,
+ 1924.0,
+ 1189.0,
+ 1924.0,
+ 1189.0,
+ 1932.0,
+ 1182.0,
+ 1932.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1196.0,
+ 1924.0,
+ 1204.0,
+ 1924.0,
+ 1204.0,
+ 1932.0,
+ 1196.0,
+ 1932.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 437.0,
+ 616.0,
+ 828.0,
+ 616.0,
+ 828.0,
+ 761.0,
+ 437.0,
+ 761.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 889.0,
+ 627.0,
+ 1245.0,
+ 627.0,
+ 1245.0,
+ 756.0,
+ 889.0,
+ 756.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1048.0,
+ 1453.0,
+ 1061.0,
+ 1453.0,
+ 1061.0,
+ 1464.0,
+ 1048.0,
+ 1464.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1074.0,
+ 1455.0,
+ 1084.0,
+ 1455.0,
+ 1084.0,
+ 1463.0,
+ 1074.0,
+ 1463.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1099.0,
+ 1454.0,
+ 1108.0,
+ 1454.0,
+ 1108.0,
+ 1464.0,
+ 1099.0,
+ 1464.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1123.0,
+ 1454.0,
+ 1132.0,
+ 1454.0,
+ 1132.0,
+ 1464.0,
+ 1123.0,
+ 1464.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1023.0,
+ 1478.0,
+ 1036.0,
+ 1478.0,
+ 1036.0,
+ 1489.0,
+ 1023.0,
+ 1489.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1074.0,
+ 1479.0,
+ 1084.0,
+ 1479.0,
+ 1084.0,
+ 1488.0,
+ 1074.0,
+ 1488.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1097.0,
+ 1478.0,
+ 1110.0,
+ 1478.0,
+ 1110.0,
+ 1489.0,
+ 1097.0,
+ 1489.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1122.0,
+ 1478.0,
+ 1134.0,
+ 1478.0,
+ 1134.0,
+ 1490.0,
+ 1122.0,
+ 1490.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1099.0,
+ 1504.0,
+ 1108.0,
+ 1504.0,
+ 1108.0,
+ 1513.0,
+ 1099.0,
+ 1513.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1124.0,
+ 1505.0,
+ 1132.0,
+ 1505.0,
+ 1132.0,
+ 1512.0,
+ 1124.0,
+ 1512.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1175.0,
+ 1502.0,
+ 1189.0,
+ 1502.0,
+ 1189.0,
+ 1514.0,
+ 1175.0,
+ 1514.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1229.0,
+ 1504.0,
+ 1237.0,
+ 1504.0,
+ 1237.0,
+ 1513.0,
+ 1229.0,
+ 1513.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1356.0,
+ 1502.0,
+ 1369.0,
+ 1502.0,
+ 1369.0,
+ 1514.0,
+ 1356.0,
+ 1514.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1050.0,
+ 1529.0,
+ 1059.0,
+ 1529.0,
+ 1059.0,
+ 1537.0,
+ 1050.0,
+ 1537.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1050.0,
+ 1553.0,
+ 1059.0,
+ 1553.0,
+ 1059.0,
+ 1561.0,
+ 1050.0,
+ 1561.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1101.0,
+ 1554.0,
+ 1106.0,
+ 1554.0,
+ 1106.0,
+ 1559.0,
+ 1101.0,
+ 1559.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1012.0,
+ 1595.0,
+ 1404.0,
+ 1595.0,
+ 1404.0,
+ 1618.0,
+ 1012.0,
+ 1618.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1009.0,
+ 1614.0,
+ 1206.0,
+ 1614.0,
+ 1206.0,
+ 1644.0,
+ 1009.0,
+ 1644.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1010.0,
+ 1972.0,
+ 1404.0,
+ 1972.0,
+ 1404.0,
+ 2001.0,
+ 1010.0,
+ 2001.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1009.0,
+ 1995.0,
+ 1157.0,
+ 1995.0,
+ 1157.0,
+ 2024.0,
+ 1009.0,
+ 2024.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 765.0,
+ 1403.0,
+ 765.0,
+ 1403.0,
+ 791.0,
+ 295.0,
+ 791.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 788.0,
+ 378.0,
+ 788.0,
+ 378.0,
+ 812.0,
+ 295.0,
+ 812.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 587.0,
+ 637.0,
+ 587.0,
+ 637.0,
+ 623.0,
+ 295.0,
+ 623.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 72.0,
+ 858.0,
+ 72.0,
+ 858.0,
+ 108.0,
+ 297.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1377.0,
+ 809.0,
+ 1377.0,
+ 809.0,
+ 1415.0,
+ 293.0,
+ 1415.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 225.0,
+ 882.0,
+ 225.0,
+ 882.0,
+ 266.0,
+ 292.0,
+ 266.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 839.0,
+ 2086.0,
+ 860.0,
+ 2086.0,
+ 860.0,
+ 2118.0,
+ 839.0,
+ 2118.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 279.0,
+ 1405.0,
+ 279.0,
+ 1405.0,
+ 312.0,
+ 296.0,
+ 312.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 306.0,
+ 1405.0,
+ 306.0,
+ 1405.0,
+ 343.0,
+ 294.0,
+ 343.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 340.0,
+ 1405.0,
+ 340.0,
+ 1405.0,
+ 373.0,
+ 295.0,
+ 373.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 369.0,
+ 1315.0,
+ 369.0,
+ 1315.0,
+ 406.0,
+ 294.0,
+ 406.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1352.0,
+ 369.0,
+ 1406.0,
+ 369.0,
+ 1406.0,
+ 406.0,
+ 1352.0,
+ 406.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 400.0,
+ 1406.0,
+ 400.0,
+ 1406.0,
+ 434.0,
+ 294.0,
+ 434.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 429.0,
+ 1406.0,
+ 429.0,
+ 1406.0,
+ 466.0,
+ 295.0,
+ 466.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 460.0,
+ 1405.0,
+ 460.0,
+ 1405.0,
+ 498.0,
+ 292.0,
+ 498.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 492.0,
+ 1403.0,
+ 492.0,
+ 1403.0,
+ 527.0,
+ 294.0,
+ 527.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 524.0,
+ 1038.0,
+ 524.0,
+ 1038.0,
+ 557.0,
+ 296.0,
+ 557.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1130.0,
+ 1407.0,
+ 1130.0,
+ 1407.0,
+ 1165.0,
+ 295.0,
+ 1165.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1161.0,
+ 1405.0,
+ 1161.0,
+ 1405.0,
+ 1196.0,
+ 295.0,
+ 1196.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1191.0,
+ 611.0,
+ 1191.0,
+ 611.0,
+ 1225.0,
+ 294.0,
+ 1225.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 703.0,
+ 1191.0,
+ 1405.0,
+ 1191.0,
+ 1405.0,
+ 1225.0,
+ 703.0,
+ 1225.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1222.0,
+ 1406.0,
+ 1222.0,
+ 1406.0,
+ 1258.0,
+ 292.0,
+ 1258.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1250.0,
+ 1406.0,
+ 1250.0,
+ 1406.0,
+ 1290.0,
+ 292.0,
+ 1290.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1282.0,
+ 1406.0,
+ 1282.0,
+ 1406.0,
+ 1320.0,
+ 292.0,
+ 1320.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1313.0,
+ 660.0,
+ 1313.0,
+ 660.0,
+ 1351.0,
+ 294.0,
+ 1351.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 821.0,
+ 1407.0,
+ 821.0,
+ 1407.0,
+ 859.0,
+ 294.0,
+ 859.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 852.0,
+ 702.0,
+ 852.0,
+ 702.0,
+ 889.0,
+ 294.0,
+ 889.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 788.0,
+ 852.0,
+ 1406.0,
+ 852.0,
+ 1406.0,
+ 889.0,
+ 788.0,
+ 889.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 884.0,
+ 1406.0,
+ 884.0,
+ 1406.0,
+ 920.0,
+ 294.0,
+ 920.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 912.0,
+ 1407.0,
+ 912.0,
+ 1407.0,
+ 953.0,
+ 292.0,
+ 953.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 945.0,
+ 1407.0,
+ 945.0,
+ 1407.0,
+ 981.0,
+ 295.0,
+ 981.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 975.0,
+ 336.0,
+ 975.0,
+ 336.0,
+ 1011.0,
+ 295.0,
+ 1011.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 422.0,
+ 975.0,
+ 1406.0,
+ 975.0,
+ 1406.0,
+ 1011.0,
+ 422.0,
+ 1011.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1005.0,
+ 474.0,
+ 1005.0,
+ 474.0,
+ 1040.0,
+ 295.0,
+ 1040.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 568.0,
+ 1005.0,
+ 1405.0,
+ 1005.0,
+ 1405.0,
+ 1040.0,
+ 568.0,
+ 1040.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1036.0,
+ 1406.0,
+ 1036.0,
+ 1406.0,
+ 1072.0,
+ 294.0,
+ 1072.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1067.0,
+ 1119.0,
+ 1067.0,
+ 1119.0,
+ 1101.0,
+ 294.0,
+ 1101.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1818.0,
+ 991.0,
+ 1818.0,
+ 991.0,
+ 1855.0,
+ 295.0,
+ 1855.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1848.0,
+ 993.0,
+ 1848.0,
+ 993.0,
+ 1884.0,
+ 294.0,
+ 1884.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1881.0,
+ 992.0,
+ 1881.0,
+ 992.0,
+ 1914.0,
+ 294.0,
+ 1914.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1911.0,
+ 992.0,
+ 1911.0,
+ 992.0,
+ 1945.0,
+ 294.0,
+ 1945.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1941.0,
+ 296.0,
+ 1941.0,
+ 296.0,
+ 1977.0,
+ 293.0,
+ 1977.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 373.0,
+ 1941.0,
+ 420.0,
+ 1941.0,
+ 420.0,
+ 1977.0,
+ 373.0,
+ 1977.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 457.0,
+ 1941.0,
+ 992.0,
+ 1941.0,
+ 992.0,
+ 1977.0,
+ 457.0,
+ 1977.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1972.0,
+ 993.0,
+ 1972.0,
+ 993.0,
+ 2007.0,
+ 294.0,
+ 2007.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 2002.0,
+ 384.0,
+ 2002.0,
+ 384.0,
+ 2038.0,
+ 292.0,
+ 2038.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1437.0,
+ 993.0,
+ 1437.0,
+ 993.0,
+ 1471.0,
+ 296.0,
+ 1471.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1467.0,
+ 992.0,
+ 1467.0,
+ 992.0,
+ 1499.0,
+ 295.0,
+ 1499.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1498.0,
+ 993.0,
+ 1498.0,
+ 993.0,
+ 1531.0,
+ 295.0,
+ 1531.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1527.0,
+ 992.0,
+ 1527.0,
+ 992.0,
+ 1562.0,
+ 295.0,
+ 1562.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1558.0,
+ 994.0,
+ 1558.0,
+ 994.0,
+ 1594.0,
+ 295.0,
+ 1594.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1587.0,
+ 478.0,
+ 1587.0,
+ 478.0,
+ 1624.0,
+ 295.0,
+ 1624.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 561.0,
+ 1587.0,
+ 995.0,
+ 1587.0,
+ 995.0,
+ 1624.0,
+ 561.0,
+ 1624.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1616.0,
+ 997.0,
+ 1616.0,
+ 997.0,
+ 1659.0,
+ 294.0,
+ 1659.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1650.0,
+ 1403.0,
+ 1650.0,
+ 1403.0,
+ 1682.0,
+ 296.0,
+ 1682.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1680.0,
+ 1406.0,
+ 1680.0,
+ 1406.0,
+ 1714.0,
+ 293.0,
+ 1714.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1712.0,
+ 1403.0,
+ 1712.0,
+ 1403.0,
+ 1745.0,
+ 296.0,
+ 1745.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1737.0,
+ 1406.0,
+ 1737.0,
+ 1406.0,
+ 1779.0,
+ 291.0,
+ 1779.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 1770.0,
+ 589.0,
+ 1770.0,
+ 589.0,
+ 1806.0,
+ 293.0,
+ 1806.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 6,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 1167,
+ 1404,
+ 1167,
+ 1404,
+ 1352,
+ 298,
+ 1352
+ ],
+ "score": 0.979
+ },
+ {
+ "category_id": 5,
+ "poly": [
+ 325,
+ 1567,
+ 1377,
+ 1567,
+ 1377,
+ 1665,
+ 325,
+ 1665
+ ],
+ "score": 0.973,
+ "html": "| Algorithm1Training Algorithm |
| 1: procedure TRAINING(Categorical Distribution D,with M nodes and N categories) Let ian integer from O to M-1 |
| 2: 34 | |
| for kpretrain steps do |
| 5: | x~D |
| 6: | c ~ Ber(σ(γ)) |
| 7: 8: | L = -log P(xlc) > Compute log-probability of data given config |
| 0slow ←Adam(0slow,VL) |
| 9: for kintervention steps do | |
| 10: | I_N←randint(O，M-1) |
| 11: | Dint := D with intervention on node I_N |
| 12: | if predicting intervention then |
| 13: | Li←O∀i |
| 14: | for Kpredict steps do |
| 15: | x~Dint |
| 16: | c ~ Ber(σ(γ)) |
| 17: 18: | Li←Li+-logPi(xlci;Oslow)∀i |
| >Accumulate NLL for every node i separately I_N←argmax(Li) |
| |
| 19: 20: | gammagrads，logregrets =[]，[] >Transfer Episode Adaptation Loop |
| 21: | for Kepisode steps do x~Dint |
| 22: | gammagrad，logregret = 0，0 |
| 23: | for kcfg steps do |
| 24: | c ~ Ber(σ(γ)) |
| 25: | Li=-log Pi(xlci;0slow）∀i |
| 26: | gammagrad += σ(γ)-c |
| 27: | logregret += ∑ Li |
| 28: | I_N |
| 29: | gammagrads.append(gammagrad) logregrets.append(logregret) |
| |
| 30: | J ←入MaxEnt LMaxEnt(γ)+ 入sparse LSparse(γ)+ 入DAG LDAG(γ) |
| 31: | Vγ←VγJ+∑gammagradskijlogregrets.softmax(0) ki |
| 32: | k γ ← Adam(γ,∀γ) |
"
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 300,
+ 1942,
+ 1402,
+ 1942,
+ 1402,
+ 2034,
+ 300,
+ 2034
+ ],
+ "score": 0.97
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 75,
+ 858,
+ 75,
+ 858,
+ 105,
+ 298,
+ 105
+ ],
+ "score": 0.903
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 299,
+ 1516,
+ 563,
+ 1516,
+ 563,
+ 1547,
+ 299,
+ 1547
+ ],
+ "score": 0.9
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 287,
+ 954,
+ 287,
+ 954,
+ 319,
+ 298,
+ 319
+ ],
+ "score": 0.866
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 836,
+ 2088,
+ 863,
+ 2088,
+ 863,
+ 2112,
+ 836,
+ 2112
+ ],
+ "score": 0.847
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 230,
+ 651,
+ 230,
+ 651,
+ 261,
+ 300,
+ 261
+ ],
+ "score": 0.833
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 352,
+ 671,
+ 352,
+ 671,
+ 385,
+ 300,
+ 385
+ ],
+ "score": 0.207
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 416,
+ 707,
+ 522,
+ 707,
+ 522,
+ 733,
+ 416,
+ 733
+ ],
+ "score": 0.89,
+ "latex": "D _ { \\mathrm { i n t } } : = D"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 491,
+ 1079,
+ 530,
+ 1079,
+ 530,
+ 1106,
+ 491,
+ 1106
+ ],
+ "score": 0.88,
+ "latex": "k _ { \\mathrm { c f g } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1354,
+ 590,
+ 1401,
+ 590,
+ 1401,
+ 617,
+ 1354,
+ 617
+ ],
+ "score": 0.79,
+ "latex": "\\theta _ { \\mathrm { s l o w } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 639,
+ 423,
+ 714,
+ 423,
+ 714,
+ 450,
+ 639,
+ 450
+ ],
+ "score": 0.77,
+ "latex": "M - 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 819,
+ 397,
+ 843,
+ 397,
+ 843,
+ 421,
+ 819,
+ 421
+ ],
+ "score": 0.77,
+ "latex": "D"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 758,
+ 969,
+ 821,
+ 969,
+ 821,
+ 996,
+ 758,
+ 996
+ ],
+ "score": 0.76,
+ "latex": "\\begin{array} { r l } { = { } } & { } \\left[ \\begin{array} { l } { \\right] } \\end{array} \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1376,
+ 818,
+ 1401,
+ 818,
+ 1401,
+ 842,
+ 1376,
+ 842
+ ],
+ "score": 0.76,
+ "latex": "D"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 765,
+ 1052,
+ 815,
+ 1052,
+ 815,
+ 1077,
+ 765,
+ 1077
+ ],
+ "score": 0.76,
+ "latex": "\\qquad = \\quad 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1294,
+ 1166,
+ 1314,
+ 1166,
+ 1314,
+ 1190,
+ 1294,
+ 1190
+ ],
+ "score": 0.75,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 904,
+ 397,
+ 934,
+ 397,
+ 934,
+ 421,
+ 904,
+ 421
+ ],
+ "score": 0.74,
+ "latex": "M"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1376,
+ 507,
+ 1401,
+ 507,
+ 1401,
+ 531,
+ 1376,
+ 531
+ ],
+ "score": 0.67,
+ "latex": "D"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1283,
+ 874,
+ 1295,
+ 874,
+ 1295,
+ 896,
+ 1283,
+ 896
+ ],
+ "score": 0.66,
+ "latex": "_ { i }"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 487,
+ 1104,
+ 810,
+ 1104,
+ 810,
+ 1243,
+ 487,
+ 1243
+ ],
+ "score": 0.65,
+ "latex": "\\begin{array} { r l r } & { \\stackrel { c \\sim \\mathrm { B e r } ( \\sigma ( \\gamma ) ) } { \\sim } } & \\\\ & { L _ { i } = - \\log P _ { i } ( x | c _ { i } ; \\theta _ { \\mathrm { s l o w } } ) \\quad \\forall i } & \\\\ & { { \\tt g a m a g r a d } \\quad + = \\quad \\sigma ( \\gamma ) - c } & \\\\ & { { \\tt l o g r e g r e t } \\quad + = \\quad \\displaystyle \\sum _ { i \\neq \\mathbb { I } _ { - } \\mathbb { N } } L _ { i } } & \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 418,
+ 681,
+ 499,
+ 681,
+ 499,
+ 705,
+ 418,
+ 705
+ ],
+ "score": 0.65,
+ "latex": "\\mathbb { I } \\mathbb { N } "
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 453,
+ 1023,
+ 547,
+ 1023,
+ 547,
+ 1051,
+ 453,
+ 1051
+ ],
+ "score": 0.63,
+ "latex": "x \\sim D _ { \\mathrm { i n t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 454,
+ 996,
+ 522,
+ 996,
+ 522,
+ 1021,
+ 454,
+ 1021
+ ],
+ "score": 0.59,
+ "latex": "k _ { \\mathrm { e p i s o d e } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 419,
+ 425,
+ 432,
+ 425,
+ 432,
+ 448,
+ 419,
+ 448
+ ],
+ "score": 0.58,
+ "latex": "_ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 415,
+ 480,
+ 484,
+ 480,
+ 484,
+ 505,
+ 415,
+ 505
+ ],
+ "score": 0.58,
+ "latex": "k _ { \\mathrm { p r e t r a i n } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1044,
+ 397,
+ 1070,
+ 397,
+ 1070,
+ 421,
+ 1044,
+ 421
+ ],
+ "score": 0.52,
+ "latex": "N"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 442,
+ 761,
+ 848,
+ 761,
+ 848,
+ 940,
+ 442,
+ 940
+ ],
+ "score": 0.47,
+ "latex": "\\begin{array} { r l } & { \\mathrm { { L } } _ { i } \\gets 0 \\quad \\forall i } \\\\ & { \\mathrm { \\textbf { f o r } } k _ { \\mathrm { p r e d i c t } } \\mathrm { \\ s t e p s \\ } \\mathbf { d o } } \\\\ & { \\qquad x \\sim D _ { \\mathrm { i n t } } } \\\\ & { \\qquad c \\sim \\mathrm { B e r } ( \\sigma ( \\gamma ) ) } \\\\ & { \\qquad L _ { i } \\gets L _ { i } + - \\log P _ { i } ( x | c _ { i } ; \\theta _ { \\mathrm { s l o w } } ) \\forall i } \\\\ & { \\qquad \\mathrm { { I } } _ { - } \\mathrm { N } \\gets \\mathrm { a r g m a x } ( L _ { i } ) } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 491,
+ 792,
+ 555,
+ 792,
+ 555,
+ 816,
+ 491,
+ 816
+ ],
+ "score": 0.47,
+ "latex": "k _ { \\mathrm { p r e d i c t } }"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 452,
+ 761,
+ 587,
+ 761,
+ 587,
+ 789,
+ 452,
+ 789
+ ],
+ "score": 0.45,
+ "latex": "L _ { i } \\gets \\dot { 0 } \\quad \\forall i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1097,
+ 1365,
+ 1120,
+ 1365,
+ 1120,
+ 1388,
+ 1097,
+ 1388
+ ],
+ "score": 0.44,
+ "latex": "k i"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 415,
+ 507,
+ 490,
+ 507,
+ 490,
+ 531,
+ 415,
+ 531
+ ],
+ "score": 0.39,
+ "latex": "x \\sim D"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 454,
+ 1023,
+ 547,
+ 1023,
+ 547,
+ 1051,
+ 454,
+ 1051
+ ],
+ "score": 0.36,
+ "latex": "x \\sim D _ { \\mathrm { i n t } }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 661,
+ 679,
+ 744,
+ 679,
+ 744,
+ 706,
+ 661,
+ 706
+ ],
+ "score": 0.36,
+ "latex": "M - 1 )"
+ },
+ {
+ "category_id": 14,
+ "poly": [
+ 411,
+ 484,
+ 699,
+ 484,
+ 699,
+ 622,
+ 411,
+ 622
+ ],
+ "score": 0.28,
+ "latex": "\\begin{array} { r l } & { k _ { \\mathrm { p r e t r a i n } } ~ \\mathrm { s t e p s } ~ \\mathbf { d o } } \\\\ & { x \\sim D } \\\\ & { c \\sim \\mathrm { B e r } ( \\sigma ( \\gamma ) ) } \\\\ & { L = - \\log P ( x | c ) } \\\\ & { \\theta _ { \\mathrm { s l o w } } \\mathrm { A d a m } ( \\theta _ { \\mathrm { s l o w } } , \\nabla _ { \\theta } L ) } \\end{array}"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 639,
+ 1162,
+ 787,
+ 1162,
+ 787,
+ 1189,
+ 639,
+ 1189
+ ],
+ "score": 0.27,
+ "latex": "+ = \\ \\sigma ( \\gamma ) - c"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 298.0,
+ 73.0,
+ 858.0,
+ 73.0,
+ 858.0,
+ 108.0,
+ 298.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1513.0,
+ 566.0,
+ 1513.0,
+ 566.0,
+ 1551.0,
+ 295.0,
+ 1551.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 832.0,
+ 2084.0,
+ 868.0,
+ 2084.0,
+ 868.0,
+ 2123.0,
+ 832.0,
+ 2123.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 230.0,
+ 653.0,
+ 230.0,
+ 653.0,
+ 263.0,
+ 297.0,
+ 263.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 346.0,
+ 671.0,
+ 346.0,
+ 671.0,
+ 393.0,
+ 294.0,
+ 393.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1556.0,
+ 1406.0,
+ 1556.0,
+ 1406.0,
+ 1594.0,
+ 294.0,
+ 1594.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1588.0,
+ 1406.0,
+ 1588.0,
+ 1406.0,
+ 1622.0,
+ 296.0,
+ 1622.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1618.0,
+ 1406.0,
+ 1618.0,
+ 1406.0,
+ 1651.0,
+ 292.0,
+ 1651.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1648.0,
+ 1405.0,
+ 1648.0,
+ 1405.0,
+ 1682.0,
+ 295.0,
+ 1682.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1680.0,
+ 1406.0,
+ 1680.0,
+ 1406.0,
+ 1714.0,
+ 295.0,
+ 1714.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1710.0,
+ 1406.0,
+ 1710.0,
+ 1406.0,
+ 1743.0,
+ 294.0,
+ 1743.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1741.0,
+ 1406.0,
+ 1741.0,
+ 1406.0,
+ 1773.0,
+ 294.0,
+ 1773.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1772.0,
+ 1405.0,
+ 1772.0,
+ 1405.0,
+ 1802.0,
+ 295.0,
+ 1802.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1800.0,
+ 1407.0,
+ 1800.0,
+ 1407.0,
+ 1836.0,
+ 292.0,
+ 1836.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1831.0,
+ 1406.0,
+ 1831.0,
+ 1406.0,
+ 1865.0,
+ 295.0,
+ 1865.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1863.0,
+ 1405.0,
+ 1863.0,
+ 1405.0,
+ 1893.0,
+ 295.0,
+ 1893.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1892.0,
+ 1307.0,
+ 1892.0,
+ 1307.0,
+ 1927.0,
+ 292.0,
+ 1927.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1940.0,
+ 1407.0,
+ 1940.0,
+ 1407.0,
+ 1981.0,
+ 294.0,
+ 1981.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1974.0,
+ 1403.0,
+ 1974.0,
+ 1403.0,
+ 2007.0,
+ 296.0,
+ 2007.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 2004.0,
+ 1406.0,
+ 2004.0,
+ 1406.0,
+ 2037.0,
+ 296.0,
+ 2037.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 285.0,
+ 955.0,
+ 285.0,
+ 955.0,
+ 322.0,
+ 295.0,
+ 322.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 13,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1394,
+ 1406,
+ 1394,
+ 1406,
+ 1610,
+ 297,
+ 1610
+ ],
+ "score": 0.98
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 299,
+ 229,
+ 1403,
+ 229,
+ 1403,
+ 352,
+ 299,
+ 352
+ ],
+ "score": 0.976
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 629,
+ 1403,
+ 629,
+ 1403,
+ 752,
+ 298,
+ 752
+ ],
+ "score": 0.974
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 767,
+ 1404,
+ 767,
+ 1404,
+ 861,
+ 298,
+ 861
+ ],
+ "score": 0.972
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 445,
+ 1671,
+ 1254,
+ 1671,
+ 1254,
+ 1976,
+ 445,
+ 1976
+ ],
+ "score": 0.968
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 314,
+ 930,
+ 1279,
+ 930,
+ 1279,
+ 1191,
+ 314,
+ 1191
+ ],
+ "score": 0.965
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 459,
+ 1400,
+ 459,
+ 1400,
+ 523,
+ 298,
+ 523
+ ],
+ "score": 0.952
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 298,
+ 1223,
+ 1402,
+ 1223,
+ 1402,
+ 1270,
+ 298,
+ 1270
+ ],
+ "score": 0.925
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 356,
+ 2001,
+ 1340,
+ 2001,
+ 1340,
+ 2028,
+ 356,
+ 2028
+ ],
+ "score": 0.923
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 301,
+ 400,
+ 643,
+ 400,
+ 643,
+ 431,
+ 301,
+ 431
+ ],
+ "score": 0.897
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 300,
+ 568,
+ 580,
+ 568,
+ 580,
+ 601,
+ 300,
+ 601
+ ],
+ "score": 0.896
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 301,
+ 1339,
+ 738,
+ 1339,
+ 738,
+ 1370,
+ 301,
+ 1370
+ ],
+ "score": 0.892
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 836,
+ 2088,
+ 863,
+ 2088,
+ 863,
+ 2112,
+ 836,
+ 2112
+ ],
+ "score": 0.87
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 75,
+ 858,
+ 75,
+ 858,
+ 104,
+ 298,
+ 104
+ ],
+ "score": 0.657
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 75,
+ 857,
+ 75,
+ 857,
+ 105,
+ 298,
+ 105
+ ],
+ "score": 0.252
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1123,
+ 692,
+ 1218,
+ 692,
+ 1218,
+ 720,
+ 1123,
+ 720
+ ],
+ "score": 0.88,
+ "latex": "B A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1016,
+ 692,
+ 1110,
+ 692,
+ 1110,
+ 720,
+ 1016,
+ 720
+ ],
+ "score": 0.88,
+ "latex": "A B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 670,
+ 496,
+ 691,
+ 496,
+ 691,
+ 523,
+ 670,
+ 523
+ ],
+ "score": 0.82,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1332,
+ 692,
+ 1358,
+ 692,
+ 1358,
+ 719,
+ 1332,
+ 719
+ ],
+ "score": 0.81,
+ "latex": "B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 906,
+ 693,
+ 930,
+ 693,
+ 930,
+ 719,
+ 906,
+ 719
+ ],
+ "score": 0.81,
+ "latex": "B"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 959,
+ 770,
+ 977,
+ 770,
+ 977,
+ 796,
+ 959,
+ 796
+ ],
+ "score": 0.81,
+ "latex": "\\theta"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1175,
+ 801,
+ 1234,
+ 801,
+ 1234,
+ 829,
+ 1175,
+ 829
+ ],
+ "score": 0.8,
+ "latex": "- 1 . 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1255,
+ 693,
+ 1279,
+ 693,
+ 1279,
+ 719,
+ 1255,
+ 719
+ ],
+ "score": 0.76,
+ "latex": "A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 828,
+ 693,
+ 852,
+ 693,
+ 852,
+ 719,
+ 828,
+ 719
+ ],
+ "score": 0.75,
+ "latex": "A"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 708,
+ 636,
+ 729,
+ 636,
+ 729,
+ 658,
+ 708,
+ 658
+ ],
+ "score": 0.73,
+ "latex": "n"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 418,
+ 636,
+ 438,
+ 636,
+ 438,
+ 658,
+ 418,
+ 658
+ ],
+ "score": 0.72,
+ "latex": "n"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 457.0,
+ 1692.0,
+ 489.0,
+ 1692.0,
+ 489.0,
+ 1709.0,
+ 457.0,
+ 1709.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 648.0,
+ 1696.0,
+ 671.0,
+ 1696.0,
+ 671.0,
+ 1707.0,
+ 648.0,
+ 1707.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 926.0,
+ 1688.0,
+ 956.0,
+ 1688.0,
+ 956.0,
+ 1705.0,
+ 926.0,
+ 1705.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1006.0,
+ 1689.0,
+ 1039.0,
+ 1689.0,
+ 1039.0,
+ 1715.0,
+ 1006.0,
+ 1715.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1166.0,
+ 1696.0,
+ 1201.0,
+ 1696.0,
+ 1201.0,
+ 1713.0,
+ 1166.0,
+ 1713.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1018.0,
+ 1755.0,
+ 1026.0,
+ 1755.0,
+ 1026.0,
+ 1762.0,
+ 1018.0,
+ 1762.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1009.0,
+ 1772.0,
+ 1034.0,
+ 1772.0,
+ 1034.0,
+ 1798.0,
+ 1009.0,
+ 1798.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 823.0,
+ 1794.0,
+ 831.0,
+ 1794.0,
+ 831.0,
+ 1802.0,
+ 823.0,
+ 1802.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 860.0,
+ 1794.0,
+ 868.0,
+ 1794.0,
+ 868.0,
+ 1802.0,
+ 860.0,
+ 1802.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 580.0,
+ 1798.0,
+ 588.0,
+ 1798.0,
+ 588.0,
+ 1807.0,
+ 580.0,
+ 1807.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 552.0,
+ 1820.0,
+ 581.0,
+ 1820.0,
+ 581.0,
+ 1833.0,
+ 552.0,
+ 1833.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 835.0,
+ 1818.0,
+ 853.0,
+ 1818.0,
+ 853.0,
+ 1828.0,
+ 835.0,
+ 1828.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1087.0,
+ 1844.0,
+ 1096.0,
+ 1844.0,
+ 1096.0,
+ 1852.0,
+ 1087.0,
+ 1852.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1094.0,
+ 1853.0,
+ 1131.0,
+ 1853.0,
+ 1131.0,
+ 1878.0,
+ 1094.0,
+ 1878.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1210.0,
+ 1856.0,
+ 1247.0,
+ 1856.0,
+ 1247.0,
+ 1880.0,
+ 1210.0,
+ 1880.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1107.0,
+ 1916.0,
+ 1115.0,
+ 1916.0,
+ 1115.0,
+ 1926.0,
+ 1107.0,
+ 1926.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1222.0,
+ 1914.0,
+ 1234.0,
+ 1914.0,
+ 1234.0,
+ 1928.0,
+ 1222.0,
+ 1928.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 463.0,
+ 1944.0,
+ 484.0,
+ 1944.0,
+ 484.0,
+ 1959.0,
+ 463.0,
+ 1959.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 647.0,
+ 1947.0,
+ 670.0,
+ 1947.0,
+ 670.0,
+ 1958.0,
+ 647.0,
+ 1958.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 739.0,
+ 1941.0,
+ 767.0,
+ 1941.0,
+ 767.0,
+ 1955.0,
+ 739.0,
+ 1955.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 927.0,
+ 1940.0,
+ 952.0,
+ 1940.0,
+ 952.0,
+ 1954.0,
+ 927.0,
+ 1954.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1211.0,
+ 1933.0,
+ 1246.0,
+ 1933.0,
+ 1246.0,
+ 1963.0,
+ 1211.0,
+ 1963.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 738.75,
+ 1687.5,
+ 767.75,
+ 1687.5,
+ 767.75,
+ 1705.0,
+ 738.75,
+ 1705.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1125.0,
+ 1773.5,
+ 1153.0,
+ 1773.5,
+ 1153.0,
+ 1794.5,
+ 1125.0,
+ 1794.5
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1098.0,
+ 1934.0,
+ 1124.0,
+ 1934.0,
+ 1124.0,
+ 1956.5,
+ 1098.0,
+ 1956.5
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 943.0,
+ 943.0,
+ 966.0,
+ 943.0,
+ 966.0,
+ 961.0,
+ 943.0,
+ 961.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1176.0,
+ 939.0,
+ 1228.0,
+ 939.0,
+ 1228.0,
+ 962.0,
+ 1176.0,
+ 962.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1173.0,
+ 954.0,
+ 1239.0,
+ 954.0,
+ 1239.0,
+ 980.0,
+ 1173.0,
+ 980.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 943.0,
+ 982.0,
+ 966.0,
+ 982.0,
+ 966.0,
+ 1000.0,
+ 943.0,
+ 1000.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1163.0,
+ 974.0,
+ 1220.0,
+ 974.0,
+ 1220.0,
+ 996.0,
+ 1163.0,
+ 996.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 922.0,
+ 991.0,
+ 950.0,
+ 991.0,
+ 950.0,
+ 1102.0,
+ 922.0,
+ 1102.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1176.0,
+ 991.0,
+ 1262.0,
+ 991.0,
+ 1262.0,
+ 1013.0,
+ 1176.0,
+ 1013.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 941.0,
+ 1017.0,
+ 968.0,
+ 1017.0,
+ 968.0,
+ 1042.0,
+ 941.0,
+ 1042.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 326.0,
+ 1038.0,
+ 344.0,
+ 1038.0,
+ 344.0,
+ 1057.0,
+ 326.0,
+ 1057.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 527.0,
+ 1066.0,
+ 542.0,
+ 1066.0,
+ 542.0,
+ 1081.0,
+ 527.0,
+ 1081.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 731.0,
+ 1067.0,
+ 741.0,
+ 1067.0,
+ 741.0,
+ 1076.0,
+ 731.0,
+ 1076.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 941.0,
+ 1059.0,
+ 967.0,
+ 1059.0,
+ 967.0,
+ 1081.0,
+ 941.0,
+ 1081.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 987.0,
+ 1063.0,
+ 1001.0,
+ 1063.0,
+ 1001.0,
+ 1085.0,
+ 987.0,
+ 1085.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 803.0,
+ 1090.0,
+ 818.0,
+ 1090.0,
+ 818.0,
+ 1103.0,
+ 803.0,
+ 1103.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 943.0,
+ 1101.0,
+ 966.0,
+ 1101.0,
+ 966.0,
+ 1119.0,
+ 943.0,
+ 1119.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 478.0,
+ 1115.0,
+ 494.0,
+ 1115.0,
+ 494.0,
+ 1129.0,
+ 478.0,
+ 1129.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 575.0,
+ 1112.0,
+ 596.0,
+ 1112.0,
+ 596.0,
+ 1134.0,
+ 575.0,
+ 1134.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 676.0,
+ 1114.0,
+ 693.0,
+ 1114.0,
+ 693.0,
+ 1132.0,
+ 676.0,
+ 1132.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 422.0,
+ 1132.0,
+ 448.0,
+ 1132.0,
+ 448.0,
+ 1160.0,
+ 422.0,
+ 1160.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 852.0,
+ 1144.0,
+ 865.0,
+ 1144.0,
+ 865.0,
+ 1152.0,
+ 852.0,
+ 1152.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 941.0,
+ 1139.0,
+ 970.0,
+ 1139.0,
+ 970.0,
+ 1160.0,
+ 941.0,
+ 1160.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 354.0,
+ 1165.0,
+ 415.0,
+ 1165.0,
+ 415.0,
+ 1188.0,
+ 354.0,
+ 1188.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 508.0,
+ 1165.0,
+ 561.0,
+ 1165.0,
+ 561.0,
+ 1188.0,
+ 508.0,
+ 1188.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 641.0,
+ 1165.0,
+ 728.0,
+ 1165.0,
+ 728.0,
+ 1188.0,
+ 641.0,
+ 1188.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 782.0,
+ 1164.0,
+ 887.0,
+ 1164.0,
+ 887.0,
+ 1189.0,
+ 782.0,
+ 1189.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 955.0,
+ 1159.0,
+ 984.0,
+ 1159.0,
+ 984.0,
+ 1180.0,
+ 955.0,
+ 1180.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1005.0,
+ 1159.0,
+ 1033.0,
+ 1159.0,
+ 1033.0,
+ 1180.0,
+ 1005.0,
+ 1180.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1056.0,
+ 1159.0,
+ 1181.0,
+ 1159.0,
+ 1181.0,
+ 1195.0,
+ 1056.0,
+ 1195.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1204.0,
+ 1159.0,
+ 1230.0,
+ 1159.0,
+ 1230.0,
+ 1179.0,
+ 1204.0,
+ 1179.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1246.0,
+ 1159.0,
+ 1281.0,
+ 1159.0,
+ 1281.0,
+ 1191.0,
+ 1246.0,
+ 1191.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 365.0,
+ 1077.0,
+ 392.0,
+ 1077.0,
+ 392.0,
+ 1107.0,
+ 365.0,
+ 1107.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1221.0,
+ 1406.0,
+ 1221.0,
+ 1406.0,
+ 1250.0,
+ 294.0,
+ 1250.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1245.0,
+ 491.0,
+ 1245.0,
+ 491.0,
+ 1271.0,
+ 294.0,
+ 1271.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 357.0,
+ 2002.0,
+ 1341.0,
+ 2002.0,
+ 1341.0,
+ 2029.0,
+ 357.0,
+ 2029.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 400.0,
+ 644.0,
+ 400.0,
+ 644.0,
+ 433.0,
+ 297.0,
+ 433.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 569.0,
+ 582.0,
+ 569.0,
+ 582.0,
+ 602.0,
+ 297.0,
+ 602.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1338.0,
+ 744.0,
+ 1338.0,
+ 744.0,
+ 1373.0,
+ 296.0,
+ 1373.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 832.0,
+ 2085.0,
+ 867.0,
+ 2085.0,
+ 867.0,
+ 2124.0,
+ 832.0,
+ 2124.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 72.0,
+ 858.0,
+ 72.0,
+ 858.0,
+ 108.0,
+ 297.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 298.0,
+ 73.0,
+ 858.0,
+ 73.0,
+ 858.0,
+ 108.0,
+ 298.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1395.0,
+ 1404.0,
+ 1395.0,
+ 1404.0,
+ 1429.0,
+ 295.0,
+ 1429.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 291.0,
+ 1423.0,
+ 1409.0,
+ 1423.0,
+ 1409.0,
+ 1464.0,
+ 291.0,
+ 1464.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1454.0,
+ 1406.0,
+ 1454.0,
+ 1406.0,
+ 1491.0,
+ 295.0,
+ 1491.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1486.0,
+ 1408.0,
+ 1486.0,
+ 1408.0,
+ 1522.0,
+ 292.0,
+ 1522.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1516.0,
+ 1407.0,
+ 1516.0,
+ 1407.0,
+ 1554.0,
+ 294.0,
+ 1554.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1548.0,
+ 1404.0,
+ 1548.0,
+ 1404.0,
+ 1582.0,
+ 294.0,
+ 1582.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1577.0,
+ 1203.0,
+ 1577.0,
+ 1203.0,
+ 1613.0,
+ 294.0,
+ 1613.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 229.0,
+ 1405.0,
+ 229.0,
+ 1405.0,
+ 265.0,
+ 294.0,
+ 265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 260.0,
+ 1407.0,
+ 260.0,
+ 1407.0,
+ 296.0,
+ 293.0,
+ 296.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 293.0,
+ 1404.0,
+ 293.0,
+ 1404.0,
+ 325.0,
+ 295.0,
+ 325.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 322.0,
+ 764.0,
+ 322.0,
+ 764.0,
+ 354.0,
+ 295.0,
+ 354.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 630.0,
+ 417.0,
+ 630.0,
+ 417.0,
+ 663.0,
+ 294.0,
+ 663.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 439.0,
+ 630.0,
+ 707.0,
+ 630.0,
+ 707.0,
+ 663.0,
+ 439.0,
+ 663.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 730.0,
+ 630.0,
+ 1404.0,
+ 630.0,
+ 1404.0,
+ 663.0,
+ 730.0,
+ 663.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 658.0,
+ 1407.0,
+ 658.0,
+ 1407.0,
+ 695.0,
+ 292.0,
+ 695.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 691.0,
+ 827.0,
+ 691.0,
+ 827.0,
+ 727.0,
+ 293.0,
+ 727.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 853.0,
+ 691.0,
+ 905.0,
+ 691.0,
+ 905.0,
+ 727.0,
+ 853.0,
+ 727.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 931.0,
+ 691.0,
+ 1015.0,
+ 691.0,
+ 1015.0,
+ 727.0,
+ 931.0,
+ 727.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1111.0,
+ 691.0,
+ 1122.0,
+ 691.0,
+ 1122.0,
+ 727.0,
+ 1111.0,
+ 727.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1219.0,
+ 691.0,
+ 1254.0,
+ 691.0,
+ 1254.0,
+ 727.0,
+ 1219.0,
+ 727.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1280.0,
+ 691.0,
+ 1331.0,
+ 691.0,
+ 1331.0,
+ 727.0,
+ 1280.0,
+ 727.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1359.0,
+ 691.0,
+ 1405.0,
+ 691.0,
+ 1405.0,
+ 727.0,
+ 1359.0,
+ 727.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 722.0,
+ 446.0,
+ 722.0,
+ 446.0,
+ 756.0,
+ 294.0,
+ 756.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 765.0,
+ 958.0,
+ 765.0,
+ 958.0,
+ 804.0,
+ 292.0,
+ 804.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 978.0,
+ 765.0,
+ 1405.0,
+ 765.0,
+ 1405.0,
+ 804.0,
+ 978.0,
+ 804.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 799.0,
+ 1174.0,
+ 799.0,
+ 1174.0,
+ 832.0,
+ 294.0,
+ 832.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1235.0,
+ 799.0,
+ 1405.0,
+ 799.0,
+ 1405.0,
+ 832.0,
+ 1235.0,
+ 832.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 829.0,
+ 1112.0,
+ 829.0,
+ 1112.0,
+ 864.0,
+ 293.0,
+ 864.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 456.0,
+ 1404.0,
+ 456.0,
+ 1404.0,
+ 496.0,
+ 294.0,
+ 496.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 488.0,
+ 669.0,
+ 488.0,
+ 669.0,
+ 528.0,
+ 294.0,
+ 528.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 692.0,
+ 488.0,
+ 1223.0,
+ 488.0,
+ 1223.0,
+ 528.0,
+ 692.0,
+ 528.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 14,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 559,
+ 1404,
+ 559,
+ 1404,
+ 744,
+ 298,
+ 744
+ ],
+ "score": 0.979
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 1392,
+ 1405,
+ 1392,
+ 1405,
+ 1576,
+ 297,
+ 1576
+ ],
+ "score": 0.978
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 839,
+ 1405,
+ 839,
+ 1405,
+ 960,
+ 298,
+ 960
+ ],
+ "score": 0.978
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 297,
+ 422,
+ 1404,
+ 422,
+ 1404,
+ 545,
+ 297,
+ 545
+ ],
+ "score": 0.975
+ },
+ {
+ "category_id": 1,
+ "poly": [
+ 298,
+ 282,
+ 1404,
+ 282,
+ 1404,
+ 407,
+ 298,
+ 407
+ ],
+ "score": 0.974
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 345,
+ 1010,
+ 1335,
+ 1010,
+ 1335,
+ 1240,
+ 345,
+ 1240
+ ],
+ "score": 0.969
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 626,
+ 1638,
+ 1045,
+ 1638,
+ 1045,
+ 1951,
+ 626,
+ 1951
+ ],
+ "score": 0.96
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 551,
+ 1261,
+ 1149,
+ 1261,
+ 1149,
+ 1286,
+ 551,
+ 1286
+ ],
+ "score": 0.928
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 370,
+ 1973,
+ 1321,
+ 1973,
+ 1321,
+ 1999,
+ 370,
+ 1999
+ ],
+ "score": 0.927
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 301,
+ 1335,
+ 686,
+ 1335,
+ 686,
+ 1365,
+ 301,
+ 1365
+ ],
+ "score": 0.92
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 301,
+ 780,
+ 623,
+ 780,
+ 623,
+ 810,
+ 301,
+ 810
+ ],
+ "score": 0.912
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 298,
+ 75,
+ 858,
+ 75,
+ 858,
+ 105,
+ 298,
+ 105
+ ],
+ "score": 0.897
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 836,
+ 2089,
+ 863,
+ 2089,
+ 863,
+ 2112,
+ 836,
+ 2112
+ ],
+ "score": 0.865
+ },
+ {
+ "category_id": 0,
+ "poly": [
+ 301,
+ 230,
+ 930,
+ 230,
+ 930,
+ 261,
+ 301,
+ 261
+ ],
+ "score": 0.74
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1136,
+ 458,
+ 1163,
+ 458,
+ 1163,
+ 485,
+ 1136,
+ 485
+ ],
+ "score": 0.86,
+ "latex": "\\gamma _ { i }"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 859,
+ 844,
+ 880,
+ 844,
+ 880,
+ 869,
+ 859,
+ 869
+ ],
+ "score": 0.84,
+ "latex": "\\gamma"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 399,
+ 839,
+ 434,
+ 839,
+ 434,
+ 866,
+ 399,
+ 866
+ ],
+ "score": 0.71,
+ "latex": "L 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 983,
+ 869,
+ 1020,
+ 869,
+ 1020,
+ 896,
+ 983,
+ 896
+ ],
+ "score": 0.68,
+ "latex": "L 1"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 1012,
+ 515,
+ 1078,
+ 515,
+ 1078,
+ 543,
+ 1012,
+ 543
+ ],
+ "score": 0.68,
+ "latex": "- 1 0 0"
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 459,
+ 899,
+ 495,
+ 899,
+ 495,
+ 926,
+ 459,
+ 926
+ ],
+ "score": 0.67,
+ "latex": "L 1"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 361.0,
+ 1018.0,
+ 382.0,
+ 1018.0,
+ 382.0,
+ 1036.0,
+ 361.0,
+ 1036.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 559.0,
+ 1019.0,
+ 641.0,
+ 1019.0,
+ 641.0,
+ 1039.0,
+ 559.0,
+ 1039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 701.0,
+ 1015.0,
+ 723.0,
+ 1015.0,
+ 723.0,
+ 1033.0,
+ 701.0,
+ 1033.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 899.0,
+ 1019.0,
+ 979.0,
+ 1019.0,
+ 979.0,
+ 1039.0,
+ 899.0,
+ 1039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1040.0,
+ 1017.0,
+ 1063.0,
+ 1017.0,
+ 1063.0,
+ 1034.0,
+ 1040.0,
+ 1034.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1238.0,
+ 1019.0,
+ 1319.0,
+ 1019.0,
+ 1319.0,
+ 1039.0,
+ 1238.0,
+ 1039.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 561.0,
+ 1032.0,
+ 639.0,
+ 1032.0,
+ 639.0,
+ 1055.0,
+ 561.0,
+ 1055.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 899.0,
+ 1033.0,
+ 980.0,
+ 1033.0,
+ 980.0,
+ 1056.0,
+ 899.0,
+ 1056.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1240.0,
+ 1033.0,
+ 1319.0,
+ 1033.0,
+ 1319.0,
+ 1056.0,
+ 1240.0,
+ 1056.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 346.0,
+ 1053.0,
+ 384.0,
+ 1053.0,
+ 384.0,
+ 1161.0,
+ 346.0,
+ 1161.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 560.0,
+ 1049.0,
+ 634.0,
+ 1049.0,
+ 634.0,
+ 1072.0,
+ 560.0,
+ 1072.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 683.0,
+ 1051.0,
+ 724.0,
+ 1051.0,
+ 724.0,
+ 1160.0,
+ 683.0,
+ 1160.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 899.0,
+ 1049.0,
+ 974.0,
+ 1049.0,
+ 974.0,
+ 1072.0,
+ 899.0,
+ 1072.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1016.0,
+ 1054.0,
+ 1075.0,
+ 1054.0,
+ 1075.0,
+ 1171.0,
+ 1016.0,
+ 1171.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1239.0,
+ 1048.0,
+ 1315.0,
+ 1048.0,
+ 1315.0,
+ 1071.0,
+ 1239.0,
+ 1071.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 361.0,
+ 1123.0,
+ 384.0,
+ 1123.0,
+ 384.0,
+ 1141.0,
+ 361.0,
+ 1141.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 701.0,
+ 1122.0,
+ 723.0,
+ 1122.0,
+ 723.0,
+ 1140.0,
+ 701.0,
+ 1140.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1041.0,
+ 1128.0,
+ 1063.0,
+ 1128.0,
+ 1063.0,
+ 1147.0,
+ 1041.0,
+ 1147.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 361.0,
+ 1158.0,
+ 384.0,
+ 1158.0,
+ 384.0,
+ 1175.0,
+ 361.0,
+ 1175.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 701.0,
+ 1158.0,
+ 723.0,
+ 1158.0,
+ 723.0,
+ 1175.0,
+ 701.0,
+ 1175.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1040.0,
+ 1166.0,
+ 1063.0,
+ 1166.0,
+ 1063.0,
+ 1183.0,
+ 1040.0,
+ 1183.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 361.0,
+ 1193.0,
+ 385.0,
+ 1193.0,
+ 385.0,
+ 1211.0,
+ 361.0,
+ 1211.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 700.0,
+ 1194.0,
+ 724.0,
+ 1194.0,
+ 724.0,
+ 1212.0,
+ 700.0,
+ 1212.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 373.0,
+ 1210.0,
+ 397.0,
+ 1210.0,
+ 397.0,
+ 1228.0,
+ 373.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 416.0,
+ 1210.0,
+ 440.0,
+ 1210.0,
+ 440.0,
+ 1228.0,
+ 416.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 461.0,
+ 1210.0,
+ 482.0,
+ 1210.0,
+ 482.0,
+ 1228.0,
+ 461.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 504.0,
+ 1210.0,
+ 525.0,
+ 1210.0,
+ 525.0,
+ 1227.0,
+ 504.0,
+ 1227.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 546.0,
+ 1210.0,
+ 569.0,
+ 1210.0,
+ 569.0,
+ 1228.0,
+ 546.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 588.0,
+ 1210.0,
+ 612.0,
+ 1210.0,
+ 612.0,
+ 1228.0,
+ 588.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 620.0,
+ 1205.0,
+ 659.0,
+ 1205.0,
+ 659.0,
+ 1242.0,
+ 620.0,
+ 1242.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 713.0,
+ 1210.0,
+ 737.0,
+ 1210.0,
+ 737.0,
+ 1228.0,
+ 713.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 756.0,
+ 1210.0,
+ 780.0,
+ 1210.0,
+ 780.0,
+ 1228.0,
+ 756.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 801.0,
+ 1210.0,
+ 823.0,
+ 1210.0,
+ 823.0,
+ 1228.0,
+ 801.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 845.0,
+ 1210.0,
+ 866.0,
+ 1210.0,
+ 866.0,
+ 1227.0,
+ 845.0,
+ 1227.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 885.0,
+ 1210.0,
+ 908.0,
+ 1210.0,
+ 908.0,
+ 1228.0,
+ 885.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 929.0,
+ 1210.0,
+ 951.0,
+ 1210.0,
+ 951.0,
+ 1228.0,
+ 929.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 972.0,
+ 1210.0,
+ 994.0,
+ 1210.0,
+ 994.0,
+ 1228.0,
+ 972.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1050.0,
+ 1209.0,
+ 1113.0,
+ 1209.0,
+ 1113.0,
+ 1228.0,
+ 1050.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1114.0,
+ 1209.0,
+ 1177.0,
+ 1209.0,
+ 1177.0,
+ 1228.0,
+ 1114.0,
+ 1228.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1181.0,
+ 1210.0,
+ 1208.0,
+ 1210.0,
+ 1208.0,
+ 1227.0,
+ 1181.0,
+ 1227.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1212.0,
+ 1210.0,
+ 1240.0,
+ 1210.0,
+ 1240.0,
+ 1227.0,
+ 1212.0,
+ 1227.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1245.0,
+ 1210.0,
+ 1274.0,
+ 1210.0,
+ 1274.0,
+ 1227.0,
+ 1245.0,
+ 1227.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1275.0,
+ 1205.0,
+ 1342.0,
+ 1205.0,
+ 1342.0,
+ 1243.0,
+ 1275.0,
+ 1243.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 463.0,
+ 1221.0,
+ 564.0,
+ 1221.0,
+ 564.0,
+ 1240.0,
+ 463.0,
+ 1240.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 804.0,
+ 1221.0,
+ 904.0,
+ 1221.0,
+ 904.0,
+ 1240.0,
+ 804.0,
+ 1240.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 965.0,
+ 1220.0,
+ 990.0,
+ 1220.0,
+ 990.0,
+ 1237.0,
+ 965.0,
+ 1237.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1145.0,
+ 1221.0,
+ 1243.0,
+ 1221.0,
+ 1243.0,
+ 1240.0,
+ 1145.0,
+ 1240.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 741.0,
+ 1087.5,
+ 784.0,
+ 1087.5,
+ 784.0,
+ 1104.0,
+ 741.0,
+ 1104.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1078.0,
+ 1080.5,
+ 1132.0,
+ 1080.5,
+ 1132.0,
+ 1096.0,
+ 1078.0,
+ 1096.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 828.25,
+ 1173.0,
+ 988.25,
+ 1173.0,
+ 988.25,
+ 1206.5,
+ 828.25,
+ 1206.5
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 643.0,
+ 1645.0,
+ 674.0,
+ 1645.0,
+ 674.0,
+ 1671.0,
+ 643.0,
+ 1671.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 915.0,
+ 1648.0,
+ 1029.0,
+ 1648.0,
+ 1029.0,
+ 1676.0,
+ 915.0,
+ 1676.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 915.0,
+ 1670.0,
+ 1029.0,
+ 1670.0,
+ 1029.0,
+ 1699.0,
+ 915.0,
+ 1699.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 644.0,
+ 1696.0,
+ 674.0,
+ 1696.0,
+ 674.0,
+ 1720.0,
+ 644.0,
+ 1720.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 622.0,
+ 1711.0,
+ 658.0,
+ 1711.0,
+ 658.0,
+ 1844.0,
+ 622.0,
+ 1844.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 644.0,
+ 1747.0,
+ 673.0,
+ 1747.0,
+ 673.0,
+ 1771.0,
+ 644.0,
+ 1771.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 642.0,
+ 1795.0,
+ 674.0,
+ 1795.0,
+ 674.0,
+ 1822.0,
+ 642.0,
+ 1822.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 642.0,
+ 1846.0,
+ 676.0,
+ 1846.0,
+ 676.0,
+ 1873.0,
+ 642.0,
+ 1873.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 644.0,
+ 1898.0,
+ 676.0,
+ 1898.0,
+ 676.0,
+ 1921.0,
+ 644.0,
+ 1921.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 661.0,
+ 1913.0,
+ 742.0,
+ 1913.0,
+ 742.0,
+ 1937.0,
+ 661.0,
+ 1937.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 750.0,
+ 1914.0,
+ 784.0,
+ 1914.0,
+ 784.0,
+ 1935.0,
+ 750.0,
+ 1935.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 792.0,
+ 1912.0,
+ 918.0,
+ 1912.0,
+ 918.0,
+ 1937.0,
+ 792.0,
+ 1937.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 930.0,
+ 1914.0,
+ 961.0,
+ 1914.0,
+ 961.0,
+ 1935.0,
+ 930.0,
+ 1935.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 973.0,
+ 1914.0,
+ 1006.0,
+ 1914.0,
+ 1006.0,
+ 1936.0,
+ 973.0,
+ 1936.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1018.0,
+ 1912.0,
+ 1051.0,
+ 1912.0,
+ 1051.0,
+ 1937.0,
+ 1018.0,
+ 1937.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 788.0,
+ 1926.0,
+ 923.0,
+ 1926.0,
+ 923.0,
+ 1954.0,
+ 788.0,
+ 1954.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1008.0,
+ 1927.0,
+ 1045.0,
+ 1927.0,
+ 1045.0,
+ 1952.0,
+ 1008.0,
+ 1952.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 549.0,
+ 1260.0,
+ 1151.0,
+ 1260.0,
+ 1151.0,
+ 1288.0,
+ 549.0,
+ 1288.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 371.0,
+ 1970.0,
+ 1327.0,
+ 1970.0,
+ 1327.0,
+ 2003.0,
+ 371.0,
+ 2003.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1334.0,
+ 691.0,
+ 1334.0,
+ 691.0,
+ 1369.0,
+ 295.0,
+ 1369.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 779.0,
+ 627.0,
+ 779.0,
+ 627.0,
+ 814.0,
+ 297.0,
+ 814.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 298.0,
+ 73.0,
+ 858.0,
+ 73.0,
+ 858.0,
+ 108.0,
+ 298.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 832.0,
+ 2085.0,
+ 868.0,
+ 2085.0,
+ 868.0,
+ 2124.0,
+ 832.0,
+ 2124.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 229.0,
+ 935.0,
+ 229.0,
+ 935.0,
+ 265.0,
+ 297.0,
+ 265.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 561.0,
+ 1406.0,
+ 561.0,
+ 1406.0,
+ 593.0,
+ 297.0,
+ 593.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 593.0,
+ 1405.0,
+ 593.0,
+ 1405.0,
+ 625.0,
+ 293.0,
+ 625.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 624.0,
+ 1402.0,
+ 624.0,
+ 1402.0,
+ 655.0,
+ 294.0,
+ 655.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 650.0,
+ 1404.0,
+ 650.0,
+ 1404.0,
+ 686.0,
+ 293.0,
+ 686.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 679.0,
+ 1406.0,
+ 679.0,
+ 1406.0,
+ 718.0,
+ 295.0,
+ 718.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 713.0,
+ 1015.0,
+ 713.0,
+ 1015.0,
+ 746.0,
+ 293.0,
+ 746.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1390.0,
+ 1406.0,
+ 1390.0,
+ 1406.0,
+ 1428.0,
+ 294.0,
+ 1428.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1422.0,
+ 1407.0,
+ 1422.0,
+ 1407.0,
+ 1457.0,
+ 295.0,
+ 1457.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 1455.0,
+ 1403.0,
+ 1455.0,
+ 1403.0,
+ 1487.0,
+ 296.0,
+ 1487.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1483.0,
+ 1405.0,
+ 1483.0,
+ 1405.0,
+ 1517.0,
+ 294.0,
+ 1517.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 1512.0,
+ 1405.0,
+ 1512.0,
+ 1405.0,
+ 1550.0,
+ 294.0,
+ 1550.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 1544.0,
+ 406.0,
+ 1544.0,
+ 406.0,
+ 1582.0,
+ 292.0,
+ 1582.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 835.0,
+ 398.0,
+ 835.0,
+ 398.0,
+ 875.0,
+ 293.0,
+ 875.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 435.0,
+ 835.0,
+ 858.0,
+ 835.0,
+ 858.0,
+ 875.0,
+ 435.0,
+ 875.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 881.0,
+ 835.0,
+ 1403.0,
+ 835.0,
+ 1403.0,
+ 875.0,
+ 881.0,
+ 875.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 296.0,
+ 868.0,
+ 982.0,
+ 868.0,
+ 982.0,
+ 902.0,
+ 296.0,
+ 902.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1021.0,
+ 868.0,
+ 1403.0,
+ 868.0,
+ 1403.0,
+ 902.0,
+ 1021.0,
+ 902.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 895.0,
+ 458.0,
+ 895.0,
+ 458.0,
+ 935.0,
+ 293.0,
+ 935.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 496.0,
+ 895.0,
+ 1405.0,
+ 895.0,
+ 1405.0,
+ 935.0,
+ 496.0,
+ 935.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 929.0,
+ 1214.0,
+ 929.0,
+ 1214.0,
+ 963.0,
+ 295.0,
+ 963.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 423.0,
+ 1404.0,
+ 423.0,
+ 1404.0,
+ 456.0,
+ 294.0,
+ 456.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 451.0,
+ 1135.0,
+ 451.0,
+ 1135.0,
+ 488.0,
+ 292.0,
+ 488.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1164.0,
+ 451.0,
+ 1405.0,
+ 451.0,
+ 1405.0,
+ 488.0,
+ 1164.0,
+ 488.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 485.0,
+ 1404.0,
+ 485.0,
+ 1404.0,
+ 518.0,
+ 292.0,
+ 518.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 513.0,
+ 1011.0,
+ 513.0,
+ 1011.0,
+ 547.0,
+ 295.0,
+ 547.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1079.0,
+ 513.0,
+ 1087.0,
+ 513.0,
+ 1087.0,
+ 547.0,
+ 1079.0,
+ 547.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 282.0,
+ 1405.0,
+ 282.0,
+ 1405.0,
+ 319.0,
+ 294.0,
+ 319.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 313.0,
+ 1409.0,
+ 313.0,
+ 1409.0,
+ 348.0,
+ 292.0,
+ 348.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 342.0,
+ 1405.0,
+ 342.0,
+ 1405.0,
+ 381.0,
+ 292.0,
+ 381.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 293.0,
+ 374.0,
+ 1169.0,
+ 374.0,
+ 1169.0,
+ 410.0,
+ 293.0,
+ 410.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 15,
+ "width": 1700,
+ "height": 2200
+ }
+ },
+ {
+ "layout_dets": [
+ {
+ "category_id": 3,
+ "poly": [
+ 417,
+ 1334,
+ 1304,
+ 1334,
+ 1304,
+ 1734,
+ 417,
+ 1734
+ ],
+ "score": 0.971
+ },
+ {
+ "category_id": 3,
+ "poly": [
+ 388,
+ 470,
+ 1298,
+ 470,
+ 1298,
+ 800,
+ 388,
+ 800
+ ],
+ "score": 0.961
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 293,
+ 830,
+ 1406,
+ 830,
+ 1406,
+ 876,
+ 293,
+ 876
+ ],
+ "score": 0.944
+ },
+ {
+ "category_id": 4,
+ "poly": [
+ 288,
+ 1760,
+ 1403,
+ 1760,
+ 1403,
+ 1809,
+ 288,
+ 1809
+ ],
+ "score": 0.926
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 297,
+ 76,
+ 858,
+ 76,
+ 858,
+ 104,
+ 297,
+ 104
+ ],
+ "score": 0.886
+ },
+ {
+ "category_id": 2,
+ "poly": [
+ 835,
+ 2088,
+ 863,
+ 2088,
+ 863,
+ 2112,
+ 835,
+ 2112
+ ],
+ "score": 0.847
+ },
+ {
+ "category_id": 13,
+ "poly": [
+ 443,
+ 1362,
+ 455,
+ 1362,
+ 455,
+ 1375,
+ 443,
+ 1375
+ ],
+ "score": 0.4,
+ "latex": "\\circ"
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 785.0,
+ 1335.0,
+ 820.0,
+ 1335.0,
+ 820.0,
+ 1357.0,
+ 785.0,
+ 1357.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1130.0,
+ 1377.0,
+ 1247.0,
+ 1377.0,
+ 1247.0,
+ 1398.0,
+ 1130.0,
+ 1398.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 785.0,
+ 1392.0,
+ 816.0,
+ 1392.0,
+ 816.0,
+ 1415.0,
+ 785.0,
+ 1415.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1119.0,
+ 1392.0,
+ 1265.0,
+ 1392.0,
+ 1265.0,
+ 1415.0,
+ 1119.0,
+ 1415.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 760.0,
+ 1399.0,
+ 790.0,
+ 1399.0,
+ 790.0,
+ 1642.0,
+ 760.0,
+ 1642.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1133.0,
+ 1411.0,
+ 1247.0,
+ 1411.0,
+ 1247.0,
+ 1430.0,
+ 1133.0,
+ 1430.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1133.0,
+ 1428.0,
+ 1263.0,
+ 1428.0,
+ 1263.0,
+ 1444.0,
+ 1133.0,
+ 1444.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 785.0,
+ 1448.0,
+ 816.0,
+ 1448.0,
+ 816.0,
+ 1475.0,
+ 785.0,
+ 1475.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 592.0,
+ 1499.0,
+ 602.0,
+ 1499.0,
+ 602.0,
+ 1508.0,
+ 592.0,
+ 1508.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 784.0,
+ 1509.0,
+ 814.0,
+ 1509.0,
+ 814.0,
+ 1533.0,
+ 784.0,
+ 1533.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 563.0,
+ 1523.0,
+ 584.0,
+ 1523.0,
+ 584.0,
+ 1549.0,
+ 563.0,
+ 1549.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 784.0,
+ 1567.0,
+ 814.0,
+ 1567.0,
+ 814.0,
+ 1591.0,
+ 784.0,
+ 1591.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 784.0,
+ 1628.0,
+ 815.0,
+ 1628.0,
+ 815.0,
+ 1651.0,
+ 784.0,
+ 1651.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 438.0,
+ 1691.0,
+ 459.0,
+ 1691.0,
+ 459.0,
+ 1716.0,
+ 438.0,
+ 1716.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 689.0,
+ 1693.0,
+ 708.0,
+ 1693.0,
+ 708.0,
+ 1713.0,
+ 689.0,
+ 1713.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 784.0,
+ 1685.0,
+ 832.0,
+ 1685.0,
+ 832.0,
+ 1720.0,
+ 784.0,
+ 1720.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 943.0,
+ 1696.0,
+ 975.0,
+ 1696.0,
+ 975.0,
+ 1720.0,
+ 943.0,
+ 1720.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1082.0,
+ 1695.0,
+ 1123.0,
+ 1695.0,
+ 1123.0,
+ 1722.0,
+ 1082.0,
+ 1722.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1225.0,
+ 1695.0,
+ 1266.0,
+ 1695.0,
+ 1266.0,
+ 1722.0,
+ 1225.0,
+ 1722.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1015.0,
+ 1709.0,
+ 1098.0,
+ 1709.0,
+ 1098.0,
+ 1743.0,
+ 1015.0,
+ 1743.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 403.0,
+ 470.0,
+ 428.0,
+ 470.0,
+ 428.0,
+ 489.0,
+ 403.0,
+ 489.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 890.0,
+ 470.0,
+ 916.0,
+ 470.0,
+ 916.0,
+ 489.0,
+ 890.0,
+ 489.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 715.0,
+ 486.0,
+ 794.0,
+ 486.0,
+ 794.0,
+ 507.0,
+ 715.0,
+ 507.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1220.0,
+ 485.0,
+ 1282.0,
+ 485.0,
+ 1282.0,
+ 509.0,
+ 1220.0,
+ 509.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 403.0,
+ 518.0,
+ 430.0,
+ 518.0,
+ 430.0,
+ 538.0,
+ 403.0,
+ 538.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 887.0,
+ 526.0,
+ 918.0,
+ 526.0,
+ 918.0,
+ 549.0,
+ 887.0,
+ 549.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 386.0,
+ 567.0,
+ 428.0,
+ 567.0,
+ 428.0,
+ 679.0,
+ 386.0,
+ 679.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 871.0,
+ 571.0,
+ 916.0,
+ 571.0,
+ 916.0,
+ 679.0,
+ 871.0,
+ 679.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 402.0,
+ 613.0,
+ 428.0,
+ 613.0,
+ 428.0,
+ 634.0,
+ 402.0,
+ 634.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 889.0,
+ 642.0,
+ 916.0,
+ 642.0,
+ 916.0,
+ 663.0,
+ 889.0,
+ 663.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 400.0,
+ 662.0,
+ 428.0,
+ 662.0,
+ 428.0,
+ 682.0,
+ 400.0,
+ 682.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 889.0,
+ 699.0,
+ 919.0,
+ 699.0,
+ 919.0,
+ 722.0,
+ 889.0,
+ 722.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 401.0,
+ 709.0,
+ 432.0,
+ 709.0,
+ 432.0,
+ 732.0,
+ 401.0,
+ 732.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 471.0,
+ 768.0,
+ 506.0,
+ 768.0,
+ 506.0,
+ 787.0,
+ 471.0,
+ 787.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 529.0,
+ 767.0,
+ 573.0,
+ 767.0,
+ 573.0,
+ 788.0,
+ 529.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 592.0,
+ 766.0,
+ 635.0,
+ 766.0,
+ 635.0,
+ 788.0,
+ 592.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 654.0,
+ 767.0,
+ 697.0,
+ 767.0,
+ 697.0,
+ 788.0,
+ 654.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 715.0,
+ 767.0,
+ 758.0,
+ 767.0,
+ 758.0,
+ 788.0,
+ 715.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 777.0,
+ 767.0,
+ 820.0,
+ 767.0,
+ 820.0,
+ 788.0,
+ 777.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 889.0,
+ 758.0,
+ 921.0,
+ 758.0,
+ 921.0,
+ 787.0,
+ 889.0,
+ 787.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 958.0,
+ 768.0,
+ 992.0,
+ 768.0,
+ 992.0,
+ 787.0,
+ 958.0,
+ 787.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1017.0,
+ 767.0,
+ 1060.0,
+ 767.0,
+ 1060.0,
+ 788.0,
+ 1017.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1079.0,
+ 767.0,
+ 1122.0,
+ 767.0,
+ 1122.0,
+ 788.0,
+ 1079.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1141.0,
+ 767.0,
+ 1183.0,
+ 767.0,
+ 1183.0,
+ 788.0,
+ 1141.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1202.0,
+ 767.0,
+ 1246.0,
+ 767.0,
+ 1246.0,
+ 788.0,
+ 1202.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1264.0,
+ 767.0,
+ 1305.0,
+ 767.0,
+ 1305.0,
+ 788.0,
+ 1264.0,
+ 788.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 557.0,
+ 781.0,
+ 669.0,
+ 781.0,
+ 669.0,
+ 803.0,
+ 557.0,
+ 803.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 1044.0,
+ 781.0,
+ 1157.0,
+ 781.0,
+ 1157.0,
+ 803.0,
+ 1044.0,
+ 803.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 408.0,
+ 757.0,
+ 430.0,
+ 757.0,
+ 430.0,
+ 783.5,
+ 408.0,
+ 783.5
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 294.0,
+ 827.0,
+ 1407.0,
+ 827.0,
+ 1407.0,
+ 859.0,
+ 294.0,
+ 859.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 292.0,
+ 850.0,
+ 564.0,
+ 850.0,
+ 564.0,
+ 879.0,
+ 292.0,
+ 879.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1760.0,
+ 1405.0,
+ 1760.0,
+ 1405.0,
+ 1789.0,
+ 295.0,
+ 1789.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 295.0,
+ 1784.0,
+ 1073.0,
+ 1784.0,
+ 1073.0,
+ 1810.0,
+ 295.0,
+ 1810.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 297.0,
+ 73.0,
+ 858.0,
+ 73.0,
+ 858.0,
+ 108.0,
+ 297.0,
+ 108.0
+ ],
+ "score": 1.0,
+ "text": ""
+ },
+ {
+ "category_id": 15,
+ "poly": [
+ 831.0,
+ 2085.0,
+ 868.0,
+ 2085.0,
+ 868.0,
+ 2124.0,
+ 831.0,
+ 2124.0
+ ],
+ "score": 1.0,
+ "text": ""
+ }
+ ],
+ "page_info": {
+ "page_no": 16,
+ "width": 1700,
+ "height": 2200
+ }
+ }
+]
\ No newline at end of file
diff --git a/parse/train/H1lS8oA5YQ/H1lS8oA5YQ_layout.pdf b/parse/train/H1lS8oA5YQ/H1lS8oA5YQ_layout.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..1bf2657106b2a2b61261dd66b24a728ddc89fbbb
--- /dev/null
+++ b/parse/train/H1lS8oA5YQ/H1lS8oA5YQ_layout.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3665e754692ab6817e4c2f2c59cb2837c6dce4f38eddb925a568c701047805b4
+size 7293243
diff --git a/parse/train/H1lS8oA5YQ/H1lS8oA5YQ_origin.pdf b/parse/train/H1lS8oA5YQ/H1lS8oA5YQ_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..097c5c39ad0093f7d8c4dd34d12040b4efec5fd6
--- /dev/null
+++ b/parse/train/H1lS8oA5YQ/H1lS8oA5YQ_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a28fa613aa60d93c8e92127494f1780b5d94fe3bb95d382deb7b163c0038cab
+size 6240847
diff --git a/parse/train/H1lS8oA5YQ/H1lS8oA5YQ_span.pdf b/parse/train/H1lS8oA5YQ/H1lS8oA5YQ_span.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..10283ed0cd2e0d603edc837f6f69daea03edc09f
--- /dev/null
+++ b/parse/train/H1lS8oA5YQ/H1lS8oA5YQ_span.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6c67f52cb282b638f994d40776f146c11ea0f446f8e011e72a294839ed34b88
+size 7300116
diff --git a/parse/train/HJxJ2h4tPr/HJxJ2h4tPr_layout.pdf b/parse/train/HJxJ2h4tPr/HJxJ2h4tPr_layout.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..00486f33faccafdc0443264ccdb299b6037bd1be
--- /dev/null
+++ b/parse/train/HJxJ2h4tPr/HJxJ2h4tPr_layout.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f481cbf08d5ab2c203651062ef3fba66edac9862be97c4ef631bfa204c5b5642
+size 2392298
diff --git a/parse/train/HJxJ2h4tPr/HJxJ2h4tPr_origin.pdf b/parse/train/HJxJ2h4tPr/HJxJ2h4tPr_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..9fd75bce5d12282ab02231a4dc4912cf104a41bc
--- /dev/null
+++ b/parse/train/HJxJ2h4tPr/HJxJ2h4tPr_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ee79596d06c8d16d7b166768fc4f5804f90f8e41af4da07003004020c557b9f
+size 2228113
diff --git a/parse/train/HJxJ2h4tPr/HJxJ2h4tPr_span.pdf b/parse/train/HJxJ2h4tPr/HJxJ2h4tPr_span.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0484af1cf60d86f62265aca6a64aa213209253e7
--- /dev/null
+++ b/parse/train/HJxJ2h4tPr/HJxJ2h4tPr_span.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0417faef656ad55ab5e61cb688d82b2608d4e48f48a0e28700f3e5f632b06837
+size 2394349
diff --git a/parse/train/Hk2aImxAb/Hk2aImxAb_layout.pdf b/parse/train/Hk2aImxAb/Hk2aImxAb_layout.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c728728571ce67a320cf698b2c9e339564c24033
--- /dev/null
+++ b/parse/train/Hk2aImxAb/Hk2aImxAb_layout.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b845fd189f4925d74e6b311d24f4227c241b3c4c1fe6e2e26c9283693a091541
+size 7835758
diff --git a/parse/train/Hk2aImxAb/Hk2aImxAb_origin.pdf b/parse/train/Hk2aImxAb/Hk2aImxAb_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..78ce738746beca7a29a4158ade5dbb10dc402dbb
--- /dev/null
+++ b/parse/train/Hk2aImxAb/Hk2aImxAb_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5b969617acb13c05116193bb427504d542189aac4771f53322d21117b5b4e59
+size 7673220
diff --git a/parse/train/Hk2aImxAb/Hk2aImxAb_span.pdf b/parse/train/Hk2aImxAb/Hk2aImxAb_span.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..6f85f1f189b2580bbe92f734fa6de0e95301caa1
--- /dev/null
+++ b/parse/train/Hk2aImxAb/Hk2aImxAb_span.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56f54662ba19f4ea577007a46e4014a8c32a25bc760202b886b8f75751cdfa0a
+size 7848492
diff --git a/parse/train/HyGySsAct7/HyGySsAct7_origin.pdf b/parse/train/HyGySsAct7/HyGySsAct7_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..9bc1a1b6a83df3e3c64f38563349357d465c9c6f
--- /dev/null
+++ b/parse/train/HyGySsAct7/HyGySsAct7_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f34a70087e7b080c054c0d2f069dd21a714d19f7042c51bb1a6e5f8ef7da8f00
+size 1136655
diff --git a/parse/train/HyGySsAct7/HyGySsAct7_span.pdf b/parse/train/HyGySsAct7/HyGySsAct7_span.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..9c8d26afb714aad802fb19d013eeb21484f6d2c4
--- /dev/null
+++ b/parse/train/HyGySsAct7/HyGySsAct7_span.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fbd57b3862688a2973586c630c3e8123f8ce00c3038df52e1b51b0a99040b83
+size 1214388
diff --git a/parse/train/HyeaSkrYPH/HyeaSkrYPH_layout.pdf b/parse/train/HyeaSkrYPH/HyeaSkrYPH_layout.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..59f812858f8e33229389666a120de348458376c8
--- /dev/null
+++ b/parse/train/HyeaSkrYPH/HyeaSkrYPH_layout.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6a73d0b137817890bed09683a7c17c4c600547768a97e4637802fa11de2c25c
+size 519284
diff --git a/parse/train/HyeaSkrYPH/HyeaSkrYPH_origin.pdf b/parse/train/HyeaSkrYPH/HyeaSkrYPH_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3dc54cff0d497873d3bb2964cfda2bfd54e0a026
--- /dev/null
+++ b/parse/train/HyeaSkrYPH/HyeaSkrYPH_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e909c525f6451757765aa355eb6685ebd735a3f444f7246dc904eed6ded303ed
+size 345372
diff --git a/parse/train/HyeaSkrYPH/HyeaSkrYPH_span.pdf b/parse/train/HyeaSkrYPH/HyeaSkrYPH_span.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..092d5c215f64333d815f1a5ab3cd035299358419
--- /dev/null
+++ b/parse/train/HyeaSkrYPH/HyeaSkrYPH_span.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41312c86fd04c442014e6ec7f05195d85fc590bc6dae0d421eb3d6dc64dc07e2
+size 521211
diff --git a/parse/train/Hyg96gBKPS/Hyg96gBKPS_layout.pdf b/parse/train/Hyg96gBKPS/Hyg96gBKPS_layout.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3145694abd055b49630da5acf36a4ec55ff2647d
--- /dev/null
+++ b/parse/train/Hyg96gBKPS/Hyg96gBKPS_layout.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f707b32bc06609442c2d1552410b93690610134f911efa892854c0afba8f27de
+size 817556
diff --git a/parse/train/Hyg96gBKPS/Hyg96gBKPS_origin.pdf b/parse/train/Hyg96gBKPS/Hyg96gBKPS_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..1c23a34b6e6d0e74bf8faeace1672f01dab702e9
--- /dev/null
+++ b/parse/train/Hyg96gBKPS/Hyg96gBKPS_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bad37e4797f2744766c95488ce1fc260e9730455320cb55a72b6292ad33e1da1
+size 575801
diff --git a/parse/train/Hyg96gBKPS/Hyg96gBKPS_span.pdf b/parse/train/Hyg96gBKPS/Hyg96gBKPS_span.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c58e3fabf1f1d8c42f1142796d30622523c6729e
--- /dev/null
+++ b/parse/train/Hyg96gBKPS/Hyg96gBKPS_span.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02adc4d99de743acd99a081fe181194289fd66f387ef1729e03cfdb1e7d20cf3
+size 815952
diff --git a/parse/train/JeweO9-QqV-/JeweO9-QqV-_middle.json b/parse/train/JeweO9-QqV-/JeweO9-QqV-_middle.json
new file mode 100644
index 0000000000000000000000000000000000000000..195aae4b0935e6d43a7bb7895cc122d78b169f1d
--- /dev/null
+++ b/parse/train/JeweO9-QqV-/JeweO9-QqV-_middle.json
@@ -0,0 +1,44487 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 78,
+ 503,
+ 115
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 505,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 505,
+ 99
+ ],
+ "score": 1.0,
+ "content": "SOGCN: SECOND-ORDER GRAPH CONVOLUTIONAL",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 98,
+ 195,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 98,
+ 195,
+ 118
+ ],
+ "score": 1.0,
+ "content": "NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 136,
+ 244,
+ 157
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 136,
+ 201,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 136,
+ 201,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Anonymous authors",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 146,
+ 245,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 146,
+ 245,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Paper under double-blind review",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 187,
+ 333,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 186,
+ 335,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 186,
+ 335,
+ 200
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 213,
+ 468,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 212,
+ 469,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 212,
+ 469,
+ 226
+ ],
+ "score": 1.0,
+ "content": "We introduce a second-order graph convolution (SoGC), a maximally localized",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 224,
+ 469,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 224,
+ 469,
+ 236
+ ],
+ "score": 1.0,
+ "content": "kernel, that can express a polynomial spectral filter with arbitrary coefficients.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 234,
+ 470,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 234,
+ 470,
+ 248
+ ],
+ "score": 1.0,
+ "content": "We contrast our SoGC with vanilla GCN, first-order (one-hop) aggregation, and",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 245,
+ 470,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 245,
+ 470,
+ 259
+ ],
+ "score": 1.0,
+ "content": "higher-order (multi-hop) aggregation by analyzing graph convolutional layers via",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 257,
+ 470,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 257,
+ 470,
+ 269
+ ],
+ "score": 1.0,
+ "content": "generalized filter space. We argue that SoGC is a simple design capable of form-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 267,
+ 468,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 267,
+ 444,
+ 280
+ ],
+ "score": 1.0,
+ "content": "ing the basic building block of graph convolution, playing the same role as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 268,
+ 468,
+ 278
+ ],
+ "score": 0.87,
+ "content": "3 \\times 3",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 278,
+ 470,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 278,
+ 470,
+ 292
+ ],
+ "score": 1.0,
+ "content": "kernels in CNNs. We build purely topological Second-Order Graph Convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 290,
+ 469,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 290,
+ 469,
+ 302
+ ],
+ "score": 1.0,
+ "content": "Networks (SoGCN) and demonstrate that SoGCN consistently achieves state-of-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 301,
+ 470,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 301,
+ 470,
+ 312
+ ],
+ "score": 1.0,
+ "content": "the-art performance on the latest benchmark. Moreover, we introduce the Gated",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 310,
+ 470,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 310,
+ 470,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Recurrent Unit (GRU) to spectral GCNs. This explorative attempt further im-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 324,
+ 273,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 324,
+ 273,
+ 335
+ ],
+ "score": 1.0,
+ "content": "proves our experimental results.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 357,
+ 206,
+ 370
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 208,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 208,
+ 373
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 383,
+ 504,
+ 439
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "Deep localized convolutional filters have achieved great success in the field of deep learning. In",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 266,
+ 407
+ ],
+ "score": 1.0,
+ "content": "image recognition, the effectiveness of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 266,
+ 395,
+ 290,
+ 405
+ ],
+ "score": 0.89,
+ "content": "3 \\times 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 291,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "kernels as the basic building block in Convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "Neural Networks (CNNs) is shown both experimentally and theoretically (Zhou, 2020). We are in-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 430
+ ],
+ "score": 1.0,
+ "content": "spired to search for the maximally localized Graph Convolution (GC) kernel with full expressiveness",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 311,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 311,
+ 441
+ ],
+ "score": 1.0,
+ "content": "power for Graph Convolutional Networks (GCNs).",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 444,
+ 505,
+ 543
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 445,
+ 505,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 445,
+ 505,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Most existing GCN methods utilize localized GCs based on one-hop aggregation scheme as the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "basic building block. Extensive works have shown performance limitations of such design due to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 479
+ ],
+ "score": 1.0,
+ "content": "over-smoothing (Li et al., 2018; Oono & Suzuki, 2019; Cai & Wang, 2020). In vanilla GCNs (Kipf",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "& Welling, 2017) the root cause of its deficiency is the lumping of the graph node self-connection",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "with pairwise neighboring connections. Recent works of Xu et al. (2019); Dehmamy et al. (2019);",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 498,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 498,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "Ming Chen et al. (2020) disentangle the effect of self-connection by adding an identity mapping",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 510,
+ 505,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 505,
+ 522
+ ],
+ "score": 1.0,
+ "content": "(so-called first-order GC). However, its lack of expressive power in filter representation remains",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 521,
+ 506,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 506,
+ 534
+ ],
+ "score": 1.0,
+ "content": "(Abu-El-Haija et al., 2019). The work of (Ming Chen et al., 2020) conjectured that the ability to",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 531,
+ 492,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 492,
+ 546
+ ],
+ "score": 1.0,
+ "content": "express a polynomial filter with arbitrary coefficients is essential for preventing over-smoothing.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 549,
+ 504,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 504,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 504,
+ 561
+ ],
+ "score": 1.0,
+ "content": "A longer propagation distance in the graph facilitates GCNs to retain its expressive power, as pointed",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "out by (Liao et al., 2019; Luan et al., 2019; Abu-El-Haija et al., 2019). The minimum propagation",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 584
+ ],
+ "score": 1.0,
+ "content": "distance needed to construct our building block of GCN remains the open question. We show that the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "minimum propagation distance is two: a two-hop graph kernel with the second-order polynomials",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 457,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 457,
+ 606
+ ],
+ "score": 1.0,
+ "content": "in adjacency matrices is sufficient. We call our graph kernel Second-Order GC (SoGC).",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 610,
+ 504,
+ 654
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "We introduce a Layer Spanning Space (LSS) framework to quantify the expressive power of multi-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 635
+ ],
+ "score": 1.0,
+ "content": "layer GCs for modeling a polynomial filter with arbitrary coefficients. By relating low-pass filtering",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "on the graph spectrum (Hoang & Maehara, 2019) with over-smoothing, one can see the lack of filter",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 485,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 485,
+ 655
+ ],
+ "score": 1.0,
+ "content": "representation power (Ming Chen et al., 2020) can lead to the performance limitation of GCN.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 660,
+ 504,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 673
+ ],
+ "score": 1.0,
+ "content": "Using the LSS framework, we show that SoGCs can approximate any linear GCNs in channel-wise",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": "filtering. Furthermore, higher-order GCs do not contribute more expressiveness, and vanilla GCN",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 694
+ ],
+ "score": 1.0,
+ "content": "or first-order GCs cannot represent all polynomial filters in general. In this sense, SoGC is the",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 693,
+ 382,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 693,
+ 382,
+ 706
+ ],
+ "score": 1.0,
+ "content": "maximally localized graph kernel with the full representation power.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 41.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 504,
+ 723
+ ],
+ "score": 1.0,
+ "content": "To validate our theory, we build Second-Order Graph Convolutional Networks (SoGCN) using",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 504,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 504,
+ 734
+ ],
+ "score": 1.0,
+ "content": "SoGC kernels. Our SoGCN using simple graph topological features consistently achieves state-",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 44.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 307,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 303,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 78,
+ 503,
+ 115
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 505,
+ 99
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 505,
+ 99
+ ],
+ "score": 1.0,
+ "content": "SOGCN: SECOND-ORDER GRAPH CONVOLUTIONAL",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 98,
+ 195,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 98,
+ 195,
+ 118
+ ],
+ "score": 1.0,
+ "content": "NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 136,
+ 244,
+ 157
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 136,
+ 201,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 136,
+ 201,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Anonymous authors",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 146,
+ 245,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 146,
+ 245,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Paper under double-blind review",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 112,
+ 136,
+ 245,
+ 159
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 187,
+ 333,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 186,
+ 335,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 186,
+ 335,
+ 200
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 213,
+ 468,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 212,
+ 469,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 212,
+ 469,
+ 226
+ ],
+ "score": 1.0,
+ "content": "We introduce a second-order graph convolution (SoGC), a maximally localized",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 224,
+ 469,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 224,
+ 469,
+ 236
+ ],
+ "score": 1.0,
+ "content": "kernel, that can express a polynomial spectral filter with arbitrary coefficients.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 234,
+ 470,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 234,
+ 470,
+ 248
+ ],
+ "score": 1.0,
+ "content": "We contrast our SoGC with vanilla GCN, first-order (one-hop) aggregation, and",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 245,
+ 470,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 245,
+ 470,
+ 259
+ ],
+ "score": 1.0,
+ "content": "higher-order (multi-hop) aggregation by analyzing graph convolutional layers via",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 141,
+ 257,
+ 470,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 257,
+ 470,
+ 269
+ ],
+ "score": 1.0,
+ "content": "generalized filter space. We argue that SoGC is a simple design capable of form-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 267,
+ 468,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 267,
+ 444,
+ 280
+ ],
+ "score": 1.0,
+ "content": "ing the basic building block of graph convolution, playing the same role as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 268,
+ 468,
+ 278
+ ],
+ "score": 0.87,
+ "content": "3 \\times 3",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 278,
+ 470,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 278,
+ 470,
+ 292
+ ],
+ "score": 1.0,
+ "content": "kernels in CNNs. We build purely topological Second-Order Graph Convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 290,
+ 469,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 290,
+ 469,
+ 302
+ ],
+ "score": 1.0,
+ "content": "Networks (SoGCN) and demonstrate that SoGCN consistently achieves state-of-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 301,
+ 470,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 301,
+ 470,
+ 312
+ ],
+ "score": 1.0,
+ "content": "the-art performance on the latest benchmark. Moreover, we introduce the Gated",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 310,
+ 470,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 310,
+ 470,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Recurrent Unit (GRU) to spectral GCNs. This explorative attempt further im-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 324,
+ 273,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 324,
+ 273,
+ 335
+ ],
+ "score": 1.0,
+ "content": "proves our experimental results.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 141,
+ 212,
+ 470,
+ 335
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 357,
+ 206,
+ 370
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 208,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 208,
+ 373
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 383,
+ 504,
+ 439
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "Deep localized convolutional filters have achieved great success in the field of deep learning. In",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 266,
+ 407
+ ],
+ "score": 1.0,
+ "content": "image recognition, the effectiveness of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 266,
+ 395,
+ 290,
+ 405
+ ],
+ "score": 0.89,
+ "content": "3 \\times 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 291,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "kernels as the basic building block in Convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "Neural Networks (CNNs) is shown both experimentally and theoretically (Zhou, 2020). We are in-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 506,
+ 430
+ ],
+ "score": 1.0,
+ "content": "spired to search for the maximally localized Graph Convolution (GC) kernel with full expressiveness",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 311,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 311,
+ 441
+ ],
+ "score": 1.0,
+ "content": "power for Graph Convolutional Networks (GCNs).",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 383,
+ 506,
+ 441
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 444,
+ 505,
+ 543
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 445,
+ 505,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 445,
+ 505,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Most existing GCN methods utilize localized GCs based on one-hop aggregation scheme as the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "basic building block. Extensive works have shown performance limitations of such design due to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 505,
+ 479
+ ],
+ "score": 1.0,
+ "content": "over-smoothing (Li et al., 2018; Oono & Suzuki, 2019; Cai & Wang, 2020). In vanilla GCNs (Kipf",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 490
+ ],
+ "score": 1.0,
+ "content": "& Welling, 2017) the root cause of its deficiency is the lumping of the graph node self-connection",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "with pairwise neighboring connections. Recent works of Xu et al. (2019); Dehmamy et al. (2019);",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 498,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 498,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "Ming Chen et al. (2020) disentangle the effect of self-connection by adding an identity mapping",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 510,
+ 505,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 505,
+ 522
+ ],
+ "score": 1.0,
+ "content": "(so-called first-order GC). However, its lack of expressive power in filter representation remains",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 521,
+ 506,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 506,
+ 534
+ ],
+ "score": 1.0,
+ "content": "(Abu-El-Haija et al., 2019). The work of (Ming Chen et al., 2020) conjectured that the ability to",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 531,
+ 492,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 492,
+ 546
+ ],
+ "score": 1.0,
+ "content": "express a polynomial filter with arbitrary coefficients is essential for preventing over-smoothing.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 105,
+ 445,
+ 506,
+ 546
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 549,
+ 504,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 504,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 504,
+ 561
+ ],
+ "score": 1.0,
+ "content": "A longer propagation distance in the graph facilitates GCNs to retain its expressive power, as pointed",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "out by (Liao et al., 2019; Luan et al., 2019; Abu-El-Haija et al., 2019). The minimum propagation",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 584
+ ],
+ "score": 1.0,
+ "content": "distance needed to construct our building block of GCN remains the open question. We show that the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "minimum propagation distance is two: a two-hop graph kernel with the second-order polynomials",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 457,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 457,
+ 606
+ ],
+ "score": 1.0,
+ "content": "in adjacency matrices is sufficient. We call our graph kernel Second-Order GC (SoGC).",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 105,
+ 549,
+ 506,
+ 606
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 610,
+ 504,
+ 654
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "We introduce a Layer Spanning Space (LSS) framework to quantify the expressive power of multi-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 620,
+ 506,
+ 635
+ ],
+ "score": 1.0,
+ "content": "layer GCs for modeling a polynomial filter with arbitrary coefficients. By relating low-pass filtering",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 632,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "on the graph spectrum (Hoang & Maehara, 2019) with over-smoothing, one can see the lack of filter",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 485,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 485,
+ 655
+ ],
+ "score": 1.0,
+ "content": "representation power (Ming Chen et al., 2020) can lead to the performance limitation of GCN.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 105,
+ 610,
+ 506,
+ 655
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 660,
+ 504,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 673
+ ],
+ "score": 1.0,
+ "content": "Using the LSS framework, we show that SoGCs can approximate any linear GCNs in channel-wise",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": "filtering. Furthermore, higher-order GCs do not contribute more expressiveness, and vanilla GCN",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 506,
+ 694
+ ],
+ "score": 1.0,
+ "content": "or first-order GCs cannot represent all polynomial filters in general. In this sense, SoGC is the",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 693,
+ 382,
+ 706
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 693,
+ 382,
+ 706
+ ],
+ "score": 1.0,
+ "content": "maximally localized graph kernel with the full representation power.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 41.5,
+ "bbox_fs": [
+ 105,
+ 660,
+ 506,
+ 706
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 504,
+ 723
+ ],
+ "score": 1.0,
+ "content": "To validate our theory, we build Second-Order Graph Convolutional Networks (SoGCN) using",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 504,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 504,
+ 734
+ ],
+ "score": 1.0,
+ "content": "SoGC kernels. Our SoGCN using simple graph topological features consistently achieves state-",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 228,
+ 505,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 505,
+ 242
+ ],
+ "score": 1.0,
+ "content": "of-the-art performance on the GNN benchmark datasets (Dwivedi et al., 2020), including citation",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 240,
+ 353,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 240,
+ 353,
+ 253
+ ],
+ "score": 1.0,
+ "content": "networks, super-pixel classification, and molecule regression.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 105,
+ 708,
+ 504,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 156
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 156
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 156
+ ],
+ "score": 0.938,
+ "type": "image",
+ "image_path": "bd912b800c8a3acce6ad8e6f4f2e219a71724593d7a9f90a3a770b0cace0ae48.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 104.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 112,
+ 104.0,
+ 498,
+ 130.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 112,
+ 130.0,
+ 498,
+ 156.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 164,
+ 505,
+ 209
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 506,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 471,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Vertex domain interpretations of vanilla GC, First-Order GC and SoGC. Denote",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 471,
+ 165,
+ 478,
+ 174
+ ],
+ "score": 0.67,
+ "content": "I",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 164,
+ 506,
+ 177
+ ],
+ "score": 1.0,
+ "content": "as the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 175,
+ 506,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 175,
+ 193,
+ 188
+ ],
+ "score": 1.0,
+ "content": "zero-hop aggregator,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 176,
+ 202,
+ 185
+ ],
+ "score": 0.53,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 175,
+ 320,
+ 188
+ ],
+ "score": 1.0,
+ "content": "the first-hop aggregator and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 175,
+ 333,
+ 185
+ ],
+ "score": 0.88,
+ "content": "A ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 175,
+ 506,
+ 188
+ ],
+ "score": 1.0,
+ "content": "the second-hop aggregator. Nodes in the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 506,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 296,
+ 199
+ ],
+ "score": 1.0,
+ "content": "same colored ring share the same weights. (a)",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 187,
+ 303,
+ 197
+ ],
+ "score": 0.64,
+ "content": "I",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 187,
+ 322,
+ 199
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 187,
+ 331,
+ 197
+ ],
+ "score": 0.68,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 187,
+ 506,
+ 199
+ ],
+ "score": 1.0,
+ "content": "of vanilla GC share the same weights. (b)",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 197,
+ 475,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 197,
+ 219,
+ 210
+ ],
+ "score": 1.0,
+ "content": "First-order GC disentangles",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 198,
+ 226,
+ 207
+ ],
+ "score": 0.63,
+ "content": "I",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 197,
+ 249,
+ 210
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 249,
+ 198,
+ 258,
+ 207
+ ],
+ "score": 0.63,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 197,
+ 413,
+ 210
+ ],
+ "score": 1.0,
+ "content": ". (c) SoGC introduces new weights for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 197,
+ 426,
+ 207
+ ],
+ "score": 0.89,
+ "content": "A ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 197,
+ 475,
+ 210
+ ],
+ "score": 1.0,
+ "content": "in addition.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 229,
+ 504,
+ 252
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 505,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 505,
+ 242
+ ],
+ "score": 1.0,
+ "content": "of-the-art performance on the GNN benchmark datasets (Dwivedi et al., 2020), including citation",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 240,
+ 353,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 240,
+ 353,
+ 253
+ ],
+ "score": 1.0,
+ "content": "networks, super-pixel classification, and molecule regression.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 257,
+ 505,
+ 335
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 505,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 505,
+ 270
+ ],
+ "score": 1.0,
+ "content": "To our best knowledge, this work is the first study that identifies the importance of the two-hop",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 280
+ ],
+ "score": 1.0,
+ "content": "neighborhood in the context of GCNs’ ability to express a polynomial filter with arbitrary coeffi-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 280,
+ 504,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 504,
+ 291
+ ],
+ "score": 1.0,
+ "content": "cients. Our model is a special but non-trivial case of Defferrard et al. (2016). Kipf & Welling (2017)",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 302
+ ],
+ "score": 1.0,
+ "content": "conducted an ablation study with GC kernels of different orders but missed the effectiveness of the",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "second-order relationships. The work of Abu-El-Haija et al. (2019) talked about muti-hop graph",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 312,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 312,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "kernels; however, they did not identify the critical importance of the second-order form. In contrast,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 323,
+ 370,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 323,
+ 370,
+ 336
+ ],
+ "score": 1.0,
+ "content": "we clarify the prominence of SoGCs in theories and experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 339,
+ 505,
+ 396
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 504,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 504,
+ 352
+ ],
+ "score": 1.0,
+ "content": "Our research on graph convolution using pure topologically relationship is orthogonal to those uses",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "geometric relations (Monti et al., 2017; Fey et al., 2018; Pei et al., 2020), or those with expressive",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "edge features (Li et al., 2016; Gilmer et al., 2017; Corso et al., 2020), and hyper-edges (Morris et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "2019; Maron et al., 2018; 2019). It is also independent with graph sampling procedures (Rong et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 284,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 284,
+ 395
+ ],
+ "score": 1.0,
+ "content": "2019; Hamilton et al., 2017; Li et al., 2019).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 411,
+ 208,
+ 423
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 209,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 209,
+ 425
+ ],
+ "score": 1.0,
+ "content": "2 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 504
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 506,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 506,
+ 448
+ ],
+ "score": 1.0,
+ "content": "We begin by reformulating spectral GCNs and introducing our notation. We are interested in a finite",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 507,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 146,
+ 461
+ ],
+ "score": 1.0,
+ "content": "graph set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 147,
+ 446,
+ 236,
+ 460
+ ],
+ "score": 0.93,
+ "content": "\\dot { \\mathcal { G } } = \\{ G _ { 1 } , \\cdot \\cdot \\cdot , \\mathbf { \\bar { G } } _ { | \\mathcal { G } | } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 446,
+ 325,
+ 461
+ ],
+ "score": 1.0,
+ "content": ". Assume each graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 447,
+ 355,
+ 457
+ ],
+ "score": 0.91,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 446,
+ 507,
+ 461
+ ],
+ "score": 1.0,
+ "content": "is simple and undirected, associated",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 458,
+ 505,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 458,
+ 198,
+ 471
+ ],
+ "score": 1.0,
+ "content": "with a finite vertex set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 459,
+ 222,
+ 470
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { V } ( G )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 222,
+ 458,
+ 273,
+ 471
+ ],
+ "score": 1.0,
+ "content": ", an edge set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 273,
+ 459,
+ 385,
+ 471
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { E } ( G ) = \\{ ( u , v ) : \\forall u v \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 458,
+ 505,
+ 471
+ ],
+ "score": 1.0,
+ "content": ", and a symmetric normalized",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 469,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 469,
+ 178,
+ 482
+ ],
+ "score": 1.0,
+ "content": "adjacency matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 470,
+ 203,
+ 481
+ ],
+ "score": 0.9,
+ "content": "A ( G )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 469,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "(Chung & Graham, 1997; Shi & Malik, 2000). Without loss of generality",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 184,
+ 494
+ ],
+ "score": 1.0,
+ "content": "and for simplicity,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 481,
+ 238,
+ 493
+ ],
+ "score": 0.92,
+ "content": "| \\mathcal { V } ( G ) | = N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 479,
+ 279,
+ 494
+ ],
+ "score": 1.0,
+ "content": "for every",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 481,
+ 309,
+ 491
+ ],
+ "score": 0.9,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 479,
+ 413,
+ 494
+ ],
+ "score": 1.0,
+ "content": ". Single-channel features",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 480,
+ 450,
+ 491
+ ],
+ "score": 0.91,
+ "content": "\\pmb { x } \\in \\mathbb { R } ^ { N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 479,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "supported in",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 491,
+ 329,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 491,
+ 131,
+ 504
+ ],
+ "score": 1.0,
+ "content": "graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 492,
+ 159,
+ 502
+ ],
+ "score": 0.9,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 491,
+ 277,
+ 504
+ ],
+ "score": 1.0,
+ "content": "is a vectorization of function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 492,
+ 324,
+ 504
+ ],
+ "score": 0.95,
+ "content": "\\mathcal { V } ( G ) \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 491,
+ 329,
+ 504
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 508,
+ 505,
+ 553
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 505,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 505,
+ 522
+ ],
+ "score": 1.0,
+ "content": "Graph Convolutions (GCs) is known as Linear Shift-Invariant (LSI) operators to adjacency matrices",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 520,
+ 505,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 520,
+ 505,
+ 532
+ ],
+ "score": 1.0,
+ "content": "(Sandryhaila & Moura, 2013). By this definition, GCs can extract features regardless of where local",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 529,
+ 506,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 529,
+ 258,
+ 544
+ ],
+ "score": 1.0,
+ "content": "structures fall. Given parameter space",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 531,
+ 287,
+ 541
+ ],
+ "score": 0.9,
+ "content": "\\Omega \\subseteq \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 288,
+ 529,
+ 464,
+ 544
+ ],
+ "score": 1.0,
+ "content": ", we write a single-channel GC (Sandryhaila",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 464,
+ 531,
+ 473,
+ 540
+ ],
+ "score": 0.44,
+ "content": "\\&",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 473,
+ 529,
+ 506,
+ 544
+ ],
+ "score": 1.0,
+ "content": "Moura,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 414,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 281,
+ 554
+ ],
+ "score": 1.0,
+ "content": "2013; Defferrard et al., 2016) as a mapping",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 541,
+ 363,
+ 553
+ ],
+ "score": 0.9,
+ "content": "f _ { \\pmb \\theta } : \\mathcal { G } \\times \\mathbb { R } ^ { N } \\mathbf { \\bar { \\mathbb { R } } } ^ { N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 541,
+ 414,
+ 554
+ ],
+ "score": 1.0,
+ "content": "such that 1:",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 247,
+ 557,
+ 363,
+ 592
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 247,
+ 557,
+ 363,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 247,
+ 557,
+ 363,
+ 592
+ ],
+ "score": 0.94,
+ "content": "f _ { \\pmb \\theta } ( G , \\pmb x ) = \\sum _ { k = 0 } ^ { K } \\theta _ { k } \\pmb { A } ( G ) ^ { k } \\pmb x ,",
+ "type": "interline_equation",
+ "image_path": "49a35a4424bd6a830c1a095bc0889a094a0f40696ee4f1bf3f2357eb14ffa918.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 32.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 247,
+ 557,
+ 363,
+ 574.5
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 247,
+ 574.5,
+ 363,
+ 592.0
+ ],
+ "spans": [],
+ "index": 33
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 595,
+ 505,
+ 633
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 594,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 594,
+ 134,
+ 612
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 595,
+ 268,
+ 610
+ ],
+ "score": 0.91,
+ "content": "\\pmb { \\theta } = \\left[ \\theta _ { 0 } \\quad \\cdot \\cdot \\cdot \\quad \\theta _ { K } \\right] ^ { T } \\in \\Omega ^ { K + 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 268,
+ 594,
+ 366,
+ 612
+ ],
+ "score": 1.0,
+ "content": "parameterizes the GC.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 598,
+ 376,
+ 608
+ ],
+ "score": 0.82,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 377,
+ 594,
+ 489,
+ 612
+ ],
+ "score": 1.0,
+ "content": "reflects the localization of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 598,
+ 501,
+ 610
+ ],
+ "score": 0.85,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 594,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": ":",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 104,
+ 608,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 608,
+ 304,
+ 624
+ ],
+ "score": 1.0,
+ "content": "a linear combination of features aggregated by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 609,
+ 334,
+ 622
+ ],
+ "score": 0.93,
+ "content": "A ( G ) ^ { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 608,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": ". Moreover, we reformulate two popular",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 399,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 399,
+ 634
+ ],
+ "score": 1.0,
+ "content": "models, vanilla GC (Figure 1a) and first-order GC (Figure 1b), as below:",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 171,
+ 636,
+ 430,
+ 651
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 171,
+ 636,
+ 430,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 171,
+ 636,
+ 430,
+ 651
+ ],
+ "score": 0.87,
+ "content": "f _ { 0 } ( G , \\pmb { x } ) = \\theta \\left( \\pmb { A } ( G ) + \\pmb { I } \\right) \\pmb { x } , \\quad f _ { 1 } ( G , \\pmb { x } ) = \\left( \\theta _ { 1 } \\pmb { A } ( G ) + \\theta _ { 0 } \\pmb { I } \\right) \\pmb { x } .",
+ "type": "interline_equation",
+ "image_path": "01e4a5ae3d592a55493826ea9b00caf5668a4d1fa50d87e8d7d4cc9ca5fcbf87.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 37,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 171,
+ 636,
+ 430,
+ 651
+ ],
+ "spans": [],
+ "index": 37
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 661,
+ 505,
+ 685
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 504,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 241,
+ 675
+ ],
+ "score": 1.0,
+ "content": "The general spectral GCNs stack",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 662,
+ 249,
+ 672
+ ],
+ "score": 0.83,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 660,
+ 487,
+ 675
+ ],
+ "score": 1.0,
+ "content": "layers of GCs (Equation 1) with nonlinear activations. Let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 488,
+ 660,
+ 504,
+ 673
+ ],
+ "score": 0.9,
+ "content": "f ^ { ( l ) }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 493,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 227,
+ 687
+ ],
+ "score": 1.0,
+ "content": "be GC layers with parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 673,
+ 313,
+ 685
+ ],
+ "score": 0.92,
+ "content": "\\pmb { \\theta } ^ { ( l ) } \\in \\Omega ^ { K + 1 } , l \\in [ L ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 671,
+ 493,
+ 687
+ ],
+ "score": 1.0,
+ "content": ", the single-channel GCNs can be written as:",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 191,
+ 689,
+ 419,
+ 705
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 191,
+ 689,
+ 419,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 191,
+ 689,
+ 419,
+ 705
+ ],
+ "score": 0.88,
+ "content": "F ( G , \\pmb { x } ) = g \\circ f ^ { ( L ) } \\circ \\sigma \\circ f ^ { ( L - 1 ) } \\circ \\cdots \\circ \\sigma \\circ f ^ { ( 1 ) } ( G , \\pmb { x } ) ,",
+ "type": "interline_equation",
+ "image_path": "cde1865d0ef5c80b278b610c60153f8d9e8c917305bce83fa7aa917061c7bc9b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 40,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 191,
+ 689,
+ 419,
+ 705
+ ],
+ "spans": [],
+ "index": 40
+ }
+ ]
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 711,
+ 507,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 709,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 709,
+ 255,
+ 723
+ ],
+ "score": 1.0,
+ "content": "1We can replace the Laplacian matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 255,
+ 712,
+ 263,
+ 721
+ ],
+ "score": 0.83,
+ "content": "\\pmb { L }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 709,
+ 495,
+ 723
+ ],
+ "score": 1.0,
+ "content": "in Defferrard et al. (2016) with the normalized adjacency matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 713,
+ 504,
+ 720
+ ],
+ "score": 0.73,
+ "content": "\\pmb { A }",
+ "type": "inline_equation"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 176,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 127,
+ 731
+ ],
+ "score": 1.0,
+ "content": "since",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 721,
+ 172,
+ 731
+ ],
+ "score": 0.91,
+ "content": "\\pmb { { \\cal L } } = \\pmb { { \\cal I } } - \\pmb { { \\cal A } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 172,
+ 721,
+ 176,
+ 731
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 307,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 156
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 156
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 156
+ ],
+ "score": 0.938,
+ "type": "image",
+ "image_path": "bd912b800c8a3acce6ad8e6f4f2e219a71724593d7a9f90a3a770b0cace0ae48.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 112,
+ 78,
+ 498,
+ 104.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 112,
+ 104.0,
+ 498,
+ 130.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 112,
+ 130.0,
+ 498,
+ 156.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 164,
+ 505,
+ 209
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 506,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 471,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Vertex domain interpretations of vanilla GC, First-Order GC and SoGC. Denote",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 471,
+ 165,
+ 478,
+ 174
+ ],
+ "score": 0.67,
+ "content": "I",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 164,
+ 506,
+ 177
+ ],
+ "score": 1.0,
+ "content": "as the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 175,
+ 506,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 175,
+ 193,
+ 188
+ ],
+ "score": 1.0,
+ "content": "zero-hop aggregator,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 176,
+ 202,
+ 185
+ ],
+ "score": 0.53,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 175,
+ 320,
+ 188
+ ],
+ "score": 1.0,
+ "content": "the first-hop aggregator and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 321,
+ 175,
+ 333,
+ 185
+ ],
+ "score": 0.88,
+ "content": "A ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 175,
+ 506,
+ 188
+ ],
+ "score": 1.0,
+ "content": "the second-hop aggregator. Nodes in the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 506,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 296,
+ 199
+ ],
+ "score": 1.0,
+ "content": "same colored ring share the same weights. (a)",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 187,
+ 303,
+ 197
+ ],
+ "score": 0.64,
+ "content": "I",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 187,
+ 322,
+ 199
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 187,
+ 331,
+ 197
+ ],
+ "score": 0.68,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 187,
+ 506,
+ 199
+ ],
+ "score": 1.0,
+ "content": "of vanilla GC share the same weights. (b)",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 197,
+ 475,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 197,
+ 219,
+ 210
+ ],
+ "score": 1.0,
+ "content": "First-order GC disentangles",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 198,
+ 226,
+ 207
+ ],
+ "score": 0.63,
+ "content": "I",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 227,
+ 197,
+ 249,
+ 210
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 249,
+ 198,
+ 258,
+ 207
+ ],
+ "score": 0.63,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 197,
+ 413,
+ 210
+ ],
+ "score": 1.0,
+ "content": ". (c) SoGC introduces new weights for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 197,
+ 426,
+ 207
+ ],
+ "score": 0.89,
+ "content": "A ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 197,
+ 475,
+ 210
+ ],
+ "score": 1.0,
+ "content": "in addition.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 229,
+ 504,
+ 252
+ ],
+ "lines": [],
+ "index": 7.5,
+ "bbox_fs": [
+ 105,
+ 228,
+ 505,
+ 253
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 257,
+ 505,
+ 335
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 505,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 256,
+ 505,
+ 270
+ ],
+ "score": 1.0,
+ "content": "To our best knowledge, this work is the first study that identifies the importance of the two-hop",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 280
+ ],
+ "score": 1.0,
+ "content": "neighborhood in the context of GCNs’ ability to express a polynomial filter with arbitrary coeffi-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 280,
+ 504,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 280,
+ 504,
+ 291
+ ],
+ "score": 1.0,
+ "content": "cients. Our model is a special but non-trivial case of Defferrard et al. (2016). Kipf & Welling (2017)",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 302
+ ],
+ "score": 1.0,
+ "content": "conducted an ablation study with GC kernels of different orders but missed the effectiveness of the",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "second-order relationships. The work of Abu-El-Haija et al. (2019) talked about muti-hop graph",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 312,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 312,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "kernels; however, they did not identify the critical importance of the second-order form. In contrast,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 323,
+ 370,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 323,
+ 370,
+ 336
+ ],
+ "score": 1.0,
+ "content": "we clarify the prominence of SoGCs in theories and experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 256,
+ 506,
+ 336
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 339,
+ 505,
+ 396
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 504,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 340,
+ 504,
+ 352
+ ],
+ "score": 1.0,
+ "content": "Our research on graph convolution using pure topologically relationship is orthogonal to those uses",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "geometric relations (Monti et al., 2017; Fey et al., 2018; Pei et al., 2020), or those with expressive",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "edge features (Li et al., 2016; Gilmer et al., 2017; Corso et al., 2020), and hyper-edges (Morris et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "2019; Maron et al., 2018; 2019). It is also independent with graph sampling procedures (Rong et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 284,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 284,
+ 395
+ ],
+ "score": 1.0,
+ "content": "2019; Hamilton et al., 2017; Li et al., 2019).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 105,
+ 340,
+ 506,
+ 395
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 411,
+ 208,
+ 423
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 209,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 209,
+ 425
+ ],
+ "score": 1.0,
+ "content": "2 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 504
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 506,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 506,
+ 448
+ ],
+ "score": 1.0,
+ "content": "We begin by reformulating spectral GCNs and introducing our notation. We are interested in a finite",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 507,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 146,
+ 461
+ ],
+ "score": 1.0,
+ "content": "graph set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 147,
+ 446,
+ 236,
+ 460
+ ],
+ "score": 0.93,
+ "content": "\\dot { \\mathcal { G } } = \\{ G _ { 1 } , \\cdot \\cdot \\cdot , \\mathbf { \\bar { G } } _ { | \\mathcal { G } | } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 446,
+ 325,
+ 461
+ ],
+ "score": 1.0,
+ "content": ". Assume each graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 447,
+ 355,
+ 457
+ ],
+ "score": 0.91,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 446,
+ 507,
+ 461
+ ],
+ "score": 1.0,
+ "content": "is simple and undirected, associated",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 458,
+ 505,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 458,
+ 198,
+ 471
+ ],
+ "score": 1.0,
+ "content": "with a finite vertex set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 459,
+ 222,
+ 470
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { V } ( G )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 222,
+ 458,
+ 273,
+ 471
+ ],
+ "score": 1.0,
+ "content": ", an edge set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 273,
+ 459,
+ 385,
+ 471
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { E } ( G ) = \\{ ( u , v ) : \\forall u v \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 458,
+ 505,
+ 471
+ ],
+ "score": 1.0,
+ "content": ", and a symmetric normalized",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 469,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 469,
+ 178,
+ 482
+ ],
+ "score": 1.0,
+ "content": "adjacency matrix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 470,
+ 203,
+ 481
+ ],
+ "score": 0.9,
+ "content": "A ( G )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 469,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "(Chung & Graham, 1997; Shi & Malik, 2000). Without loss of generality",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 184,
+ 494
+ ],
+ "score": 1.0,
+ "content": "and for simplicity,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 481,
+ 238,
+ 493
+ ],
+ "score": 0.92,
+ "content": "| \\mathcal { V } ( G ) | = N",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 479,
+ 279,
+ 494
+ ],
+ "score": 1.0,
+ "content": "for every",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 481,
+ 309,
+ 491
+ ],
+ "score": 0.9,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 479,
+ 413,
+ 494
+ ],
+ "score": 1.0,
+ "content": ". Single-channel features",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 480,
+ 450,
+ 491
+ ],
+ "score": 0.91,
+ "content": "\\pmb { x } \\in \\mathbb { R } ^ { N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 479,
+ 506,
+ 494
+ ],
+ "score": 1.0,
+ "content": "supported in",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 491,
+ 329,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 491,
+ 131,
+ 504
+ ],
+ "score": 1.0,
+ "content": "graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 492,
+ 159,
+ 502
+ ],
+ "score": 0.9,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 491,
+ 277,
+ 504
+ ],
+ "score": 1.0,
+ "content": "is a vectorization of function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 492,
+ 324,
+ 504
+ ],
+ "score": 0.95,
+ "content": "\\mathcal { V } ( G ) \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 324,
+ 491,
+ 329,
+ 504
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 435,
+ 507,
+ 504
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 508,
+ 505,
+ 553
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 505,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 508,
+ 505,
+ 522
+ ],
+ "score": 1.0,
+ "content": "Graph Convolutions (GCs) is known as Linear Shift-Invariant (LSI) operators to adjacency matrices",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 520,
+ 505,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 520,
+ 505,
+ 532
+ ],
+ "score": 1.0,
+ "content": "(Sandryhaila & Moura, 2013). By this definition, GCs can extract features regardless of where local",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 529,
+ 506,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 529,
+ 258,
+ 544
+ ],
+ "score": 1.0,
+ "content": "structures fall. Given parameter space",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 531,
+ 287,
+ 541
+ ],
+ "score": 0.9,
+ "content": "\\Omega \\subseteq \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 288,
+ 529,
+ 464,
+ 544
+ ],
+ "score": 1.0,
+ "content": ", we write a single-channel GC (Sandryhaila",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 464,
+ 531,
+ 473,
+ 540
+ ],
+ "score": 0.44,
+ "content": "\\&",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 473,
+ 529,
+ 506,
+ 544
+ ],
+ "score": 1.0,
+ "content": "Moura,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 414,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 281,
+ 554
+ ],
+ "score": 1.0,
+ "content": "2013; Defferrard et al., 2016) as a mapping",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 541,
+ 363,
+ 553
+ ],
+ "score": 0.9,
+ "content": "f _ { \\pmb \\theta } : \\mathcal { G } \\times \\mathbb { R } ^ { N } \\mathbf { \\bar { \\mathbb { R } } } ^ { N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 541,
+ 414,
+ 554
+ ],
+ "score": 1.0,
+ "content": "such that 1:",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 104,
+ 508,
+ 506,
+ 554
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 247,
+ 557,
+ 363,
+ 592
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 247,
+ 557,
+ 363,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 247,
+ 557,
+ 363,
+ 592
+ ],
+ "score": 0.94,
+ "content": "f _ { \\pmb \\theta } ( G , \\pmb x ) = \\sum _ { k = 0 } ^ { K } \\theta _ { k } \\pmb { A } ( G ) ^ { k } \\pmb x ,",
+ "type": "interline_equation",
+ "image_path": "49a35a4424bd6a830c1a095bc0889a094a0f40696ee4f1bf3f2357eb14ffa918.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 32.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 247,
+ 557,
+ 363,
+ 574.5
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 247,
+ 574.5,
+ 363,
+ 592.0
+ ],
+ "spans": [],
+ "index": 33
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 595,
+ 505,
+ 633
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 594,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 594,
+ 134,
+ 612
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 595,
+ 268,
+ 610
+ ],
+ "score": 0.91,
+ "content": "\\pmb { \\theta } = \\left[ \\theta _ { 0 } \\quad \\cdot \\cdot \\cdot \\quad \\theta _ { K } \\right] ^ { T } \\in \\Omega ^ { K + 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 268,
+ 594,
+ 366,
+ 612
+ ],
+ "score": 1.0,
+ "content": "parameterizes the GC.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 366,
+ 598,
+ 376,
+ 608
+ ],
+ "score": 0.82,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 377,
+ 594,
+ 489,
+ 612
+ ],
+ "score": 1.0,
+ "content": "reflects the localization of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 598,
+ 501,
+ 610
+ ],
+ "score": 0.85,
+ "content": "f _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 594,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": ":",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 104,
+ 608,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 608,
+ 304,
+ 624
+ ],
+ "score": 1.0,
+ "content": "a linear combination of features aggregated by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 609,
+ 334,
+ 622
+ ],
+ "score": 0.93,
+ "content": "A ( G ) ^ { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 608,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": ". Moreover, we reformulate two popular",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 399,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 399,
+ 634
+ ],
+ "score": 1.0,
+ "content": "models, vanilla GC (Figure 1a) and first-order GC (Figure 1b), as below:",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 104,
+ 594,
+ 506,
+ 634
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 171,
+ 636,
+ 430,
+ 651
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 171,
+ 636,
+ 430,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 171,
+ 636,
+ 430,
+ 651
+ ],
+ "score": 0.87,
+ "content": "f _ { 0 } ( G , \\pmb { x } ) = \\theta \\left( \\pmb { A } ( G ) + \\pmb { I } \\right) \\pmb { x } , \\quad f _ { 1 } ( G , \\pmb { x } ) = \\left( \\theta _ { 1 } \\pmb { A } ( G ) + \\theta _ { 0 } \\pmb { I } \\right) \\pmb { x } .",
+ "type": "interline_equation",
+ "image_path": "01e4a5ae3d592a55493826ea9b00caf5668a4d1fa50d87e8d7d4cc9ca5fcbf87.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 37,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 171,
+ 636,
+ 430,
+ 651
+ ],
+ "spans": [],
+ "index": 37
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 661,
+ 505,
+ 685
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 504,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 241,
+ 675
+ ],
+ "score": 1.0,
+ "content": "The general spectral GCNs stack",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 662,
+ 249,
+ 672
+ ],
+ "score": 0.83,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 250,
+ 660,
+ 487,
+ 675
+ ],
+ "score": 1.0,
+ "content": "layers of GCs (Equation 1) with nonlinear activations. Let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 488,
+ 660,
+ 504,
+ 673
+ ],
+ "score": 0.9,
+ "content": "f ^ { ( l ) }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 493,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 227,
+ 687
+ ],
+ "score": 1.0,
+ "content": "be GC layers with parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 228,
+ 673,
+ 313,
+ 685
+ ],
+ "score": 0.92,
+ "content": "\\pmb { \\theta } ^ { ( l ) } \\in \\Omega ^ { K + 1 } , l \\in [ L ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 671,
+ 493,
+ 687
+ ],
+ "score": 1.0,
+ "content": ", the single-channel GCNs can be written as:",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 105,
+ 660,
+ 504,
+ 687
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 191,
+ 689,
+ 419,
+ 705
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 191,
+ 689,
+ 419,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 191,
+ 689,
+ 419,
+ 705
+ ],
+ "score": 0.88,
+ "content": "F ( G , \\pmb { x } ) = g \\circ f ^ { ( L ) } \\circ \\sigma \\circ f ^ { ( L - 1 ) } \\circ \\cdots \\circ \\sigma \\circ f ^ { ( 1 ) } ( G , \\pmb { x } ) ,",
+ "type": "interline_equation",
+ "image_path": "cde1865d0ef5c80b278b610c60153f8d9e8c917305bce83fa7aa917061c7bc9b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 40,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 191,
+ 689,
+ 419,
+ 705
+ ],
+ "spans": [],
+ "index": 40
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 213
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 213
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 213
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "7bc99ef1a1073c3cfd8c86866fd68b0c12e398d9e7e6cf36db88f9bf2ababf8d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 123.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 123.0,
+ 504,
+ 168.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 168.0,
+ 504,
+ 213.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 222,
+ 505,
+ 277
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 222,
+ 506,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 222,
+ 506,
+ 234
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Visualizing output activation in graph spectrum domain for vanilla GCN, SoGCN, and",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "GRU variants. The test is conducted on a graph from the ZINC dataset. The spectrum is defined",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 506,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 506,
+ 257
+ ],
+ "score": 1.0,
+ "content": "as a projection of activation functions on the graph eigenvectors. SoGCN preserved higher-order",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 506,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 506,
+ 268
+ ],
+ "score": 1.0,
+ "content": "spectrum, while vanilla GCN shows over-smoothing. See Appendix F for more visualizations on",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 180,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 180,
+ 278
+ ],
+ "score": 1.0,
+ "content": "the ZINC dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 299,
+ 505,
+ 344
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 134,
+ 312
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 302,
+ 142,
+ 309
+ ],
+ "score": 0.76,
+ "content": "\\sigma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 299,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "is an element-wise activation function, the superscripts denote the corresponding layer",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 143,
+ 322
+ ],
+ "score": 1.0,
+ "content": "number,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 313,
+ 150,
+ 322
+ ],
+ "score": 0.78,
+ "content": "g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 310,
+ 453,
+ 322
+ ],
+ "score": 1.0,
+ "content": "is a task-specified readout function (e.g., softmax), the inputs are graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 453,
+ 311,
+ 486,
+ 321
+ ],
+ "score": 0.91,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 310,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 320,
+ 504,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 320,
+ 137,
+ 335
+ ],
+ "score": 1.0,
+ "content": "signals",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 137,
+ 321,
+ 172,
+ 332
+ ],
+ "score": 0.93,
+ "content": "\\pmb { x } \\in \\mathbb { R } ^ { N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 320,
+ 408,
+ 335
+ ],
+ "score": 1.0,
+ "content": ". The compositionality principle of deep learning suggests",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 409,
+ 322,
+ 417,
+ 331
+ ],
+ "score": 0.79,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 320,
+ 493,
+ 335
+ ],
+ "score": 1.0,
+ "content": "being large, while",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 322,
+ 504,
+ 331
+ ],
+ "score": 0.79,
+ "content": "K",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 296,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 296,
+ 344
+ ],
+ "score": 1.0,
+ "content": "being small and localized (LeCun et al., 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 360,
+ 395,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 358,
+ 397,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 358,
+ 397,
+ 376
+ ],
+ "score": 1.0,
+ "content": "3 OVERVIEW: SECOND-ORDER GRAPH CONVOLUTION",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 385,
+ 505,
+ 432
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "We are interested in the overall graph convolution network’s representation power of expressing a",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 101,
+ 396,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 101,
+ 396,
+ 199,
+ 429
+ ],
+ "score": 1.0,
+ "content": "polynomial filter (Equorder polynomial filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 396,
+ 490,
+ 429
+ ],
+ "score": 1.0,
+ "content": "multi-layer GCN approximate a , by stacking basic building block",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 397,
+ 500,
+ 406
+ ],
+ "score": 0.81,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 396,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "-f",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 199,
+ 408,
+ 355,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 199,
+ 408,
+ 355,
+ 422
+ ],
+ "score": 0.9,
+ "content": "\\begin{array} { r } { \\sum _ { k = 0 } ^ { K } \\theta _ { k } \\pmb { A } ( G ) ^ { k } , \\theta _ { k } \\in \\Omega , k = 0 , \\cdot \\cdot \\cdot K } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 400,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 400,
+ 433
+ ],
+ "score": 1.0,
+ "content": "graph convolution (GC) layers (Wu et al., 2019; Ming Chen et al., 2020).",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 437,
+ 505,
+ 459
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": "We formally define the second-order GC (SoGC) using the second-order polynomials of adjacency",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 146,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 146,
+ 461
+ ],
+ "score": 1.0,
+ "content": "matrices:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 214,
+ 457,
+ 396,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 214,
+ 457,
+ 396,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 214,
+ 457,
+ 396,
+ 473
+ ],
+ "score": 0.92,
+ "content": "f _ { 2 } ( G , \\pmb { x } ) = \\left( \\theta _ { 2 } \\pmb { A } ( G ) ^ { 2 } + \\theta _ { 1 } \\pmb { A } ( G ) + \\theta _ { 0 } \\pmb { I } \\right) \\pmb { x } ,",
+ "type": "interline_equation",
+ "image_path": "8f48fe7d88ac86029837946b8fcf5be90af88806cc6a4640428de062919e99db.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 214,
+ 457,
+ 396,
+ 473
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 475,
+ 505,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 134,
+ 487
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 475,
+ 212,
+ 487
+ ],
+ "score": 0.92,
+ "content": "\\theta _ { i } \\in \\mathbb { R } , i = 0 , 1 , 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 475,
+ 505,
+ 487
+ ],
+ "score": 1.0,
+ "content": "are trainable paremeters in the context of machine learning. Its vertex-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 498
+ ],
+ "score": 1.0,
+ "content": "domain interpretation is illustrated in Figure 1c. At first glance, it seems that we could stack two",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "one-hop graph convolution (GC) kernels to approximate a SoGC. However, as shown in Section 4.3,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 508,
+ 186,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 508,
+ 186,
+ 520
+ ],
+ "score": 1.0,
+ "content": "that is not the case.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 525,
+ 505,
+ 569
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "The critical insight is that graph filter approximation can be viewed as a polynomial factorization",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 548
+ ],
+ "score": 1.0,
+ "content": "problem. It is known that any univariate polynomial can be factorized into sub-polynomials of",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "degree two. Based on this fact, we show by stacking enough SoGCs (and varying their parameters)",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 319,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 319,
+ 570
+ ],
+ "score": 1.0,
+ "content": "can achieve decomposition of any polynomial filters.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 630
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "In contrast, first-order GCs are not universal approximators; two stacked one-hop GCs cannot model",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "every two-hop GC. Polynomial filter completeness of SoGC leads to better performance of GCNs.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 610
+ ],
+ "score": 1.0,
+ "content": "As shown in Figure 2, networks built with SoGC can overcome over-smoothing and extract features",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 608,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "on high-frequency bands. In the next section, we demonstrate our formal arguments on polynomial",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 619,
+ 169,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 169,
+ 632
+ ],
+ "score": 1.0,
+ "content": "approximation.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 646,
+ 314,
+ 659
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 315,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 315,
+ 661
+ ],
+ "score": 1.0,
+ "content": "4 REPRESENTATION POWER ANALYSIS",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 671,
+ 298,
+ 683
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 299,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 299,
+ 684
+ ],
+ "score": 1.0,
+ "content": "4.1 LAYER SPANNING SPACE FRAMEWORK",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 692,
+ 504,
+ 715
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 691,
+ 505,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 691,
+ 505,
+ 705
+ ],
+ "score": 1.0,
+ "content": "To illustrate the representation power of GC layers, we establish a Layer Spanning Space (LSS)",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 704,
+ 451,
+ 716
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 704,
+ 451,
+ 716
+ ],
+ "score": 1.0,
+ "content": "framework to study the graph filter space spanned by stacking multiple graph kernels.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 720,
+ 448,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 450,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 450,
+ 733
+ ],
+ "score": 1.0,
+ "content": "First, we present our mathematical devices in Definition 1, 2 with Lemma 1 as below.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 303,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 12,
+ "width": 9
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 307,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 213
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 213
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 213
+ ],
+ "score": 0.97,
+ "type": "image",
+ "image_path": "7bc99ef1a1073c3cfd8c86866fd68b0c12e398d9e7e6cf36db88f9bf2ababf8d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 504,
+ 123.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 123.0,
+ 504,
+ 168.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 168.0,
+ 504,
+ 213.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 222,
+ 505,
+ 277
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 222,
+ 506,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 222,
+ 506,
+ 234
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Visualizing output activation in graph spectrum domain for vanilla GCN, SoGCN, and",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 233,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": "GRU variants. The test is conducted on a graph from the ZINC dataset. The spectrum is defined",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 244,
+ 506,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 244,
+ 506,
+ 257
+ ],
+ "score": 1.0,
+ "content": "as a projection of activation functions on the graph eigenvectors. SoGCN preserved higher-order",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 506,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 506,
+ 268
+ ],
+ "score": 1.0,
+ "content": "spectrum, while vanilla GCN shows over-smoothing. See Appendix F for more visualizations on",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 180,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 180,
+ 278
+ ],
+ "score": 1.0,
+ "content": "the ZINC dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 299,
+ 505,
+ 344
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 134,
+ 312
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 302,
+ 142,
+ 309
+ ],
+ "score": 0.76,
+ "content": "\\sigma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 299,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "is an element-wise activation function, the superscripts denote the corresponding layer",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 143,
+ 322
+ ],
+ "score": 1.0,
+ "content": "number,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 313,
+ 150,
+ 322
+ ],
+ "score": 0.78,
+ "content": "g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 310,
+ 453,
+ 322
+ ],
+ "score": 1.0,
+ "content": "is a task-specified readout function (e.g., softmax), the inputs are graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 453,
+ 311,
+ 486,
+ 321
+ ],
+ "score": 0.91,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 486,
+ 310,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 320,
+ 504,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 320,
+ 137,
+ 335
+ ],
+ "score": 1.0,
+ "content": "signals",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 137,
+ 321,
+ 172,
+ 332
+ ],
+ "score": 0.93,
+ "content": "\\pmb { x } \\in \\mathbb { R } ^ { N }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 320,
+ 408,
+ 335
+ ],
+ "score": 1.0,
+ "content": ". The compositionality principle of deep learning suggests",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 409,
+ 322,
+ 417,
+ 331
+ ],
+ "score": 0.79,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 320,
+ 493,
+ 335
+ ],
+ "score": 1.0,
+ "content": "being large, while",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 322,
+ 504,
+ 331
+ ],
+ "score": 0.79,
+ "content": "K",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 296,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 296,
+ 344
+ ],
+ "score": 1.0,
+ "content": "being small and localized (LeCun et al., 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 104,
+ 299,
+ 506,
+ 344
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 360,
+ 395,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 358,
+ 397,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 358,
+ 397,
+ 376
+ ],
+ "score": 1.0,
+ "content": "3 OVERVIEW: SECOND-ORDER GRAPH CONVOLUTION",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 107,
+ 385,
+ 505,
+ 432
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "We are interested in the overall graph convolution network’s representation power of expressing a",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 101,
+ 396,
+ 506,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 101,
+ 396,
+ 199,
+ 429
+ ],
+ "score": 1.0,
+ "content": "polynomial filter (Equorder polynomial filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 396,
+ 490,
+ 429
+ ],
+ "score": 1.0,
+ "content": "multi-layer GCN approximate a , by stacking basic building block",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 397,
+ 500,
+ 406
+ ],
+ "score": 0.81,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 396,
+ 506,
+ 429
+ ],
+ "score": 1.0,
+ "content": "-f",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 199,
+ 408,
+ 355,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 199,
+ 408,
+ 355,
+ 422
+ ],
+ "score": 0.9,
+ "content": "\\begin{array} { r } { \\sum _ { k = 0 } ^ { K } \\theta _ { k } \\pmb { A } ( G ) ^ { k } , \\theta _ { k } \\in \\Omega , k = 0 , \\cdot \\cdot \\cdot K } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 15,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 400,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 400,
+ 433
+ ],
+ "score": 1.0,
+ "content": "graph convolution (GC) layers (Wu et al., 2019; Ming Chen et al., 2020).",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 101,
+ 384,
+ 506,
+ 433
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 437,
+ 505,
+ 459
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": "We formally define the second-order GC (SoGC) using the second-order polynomials of adjacency",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 447,
+ 146,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 447,
+ 146,
+ 461
+ ],
+ "score": 1.0,
+ "content": "matrices:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 435,
+ 505,
+ 461
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 214,
+ 457,
+ 396,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 214,
+ 457,
+ 396,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 214,
+ 457,
+ 396,
+ 473
+ ],
+ "score": 0.92,
+ "content": "f _ { 2 } ( G , \\pmb { x } ) = \\left( \\theta _ { 2 } \\pmb { A } ( G ) ^ { 2 } + \\theta _ { 1 } \\pmb { A } ( G ) + \\theta _ { 0 } \\pmb { I } \\right) \\pmb { x } ,",
+ "type": "interline_equation",
+ "image_path": "8f48fe7d88ac86029837946b8fcf5be90af88806cc6a4640428de062919e99db.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 214,
+ 457,
+ 396,
+ 473
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 475,
+ 505,
+ 520
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 134,
+ 487
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 475,
+ 212,
+ 487
+ ],
+ "score": 0.92,
+ "content": "\\theta _ { i } \\in \\mathbb { R } , i = 0 , 1 , 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 475,
+ 505,
+ 487
+ ],
+ "score": 1.0,
+ "content": "are trainable paremeters in the context of machine learning. Its vertex-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 487,
+ 505,
+ 498
+ ],
+ "score": 1.0,
+ "content": "domain interpretation is illustrated in Figure 1c. At first glance, it seems that we could stack two",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "one-hop graph convolution (GC) kernels to approximate a SoGC. However, as shown in Section 4.3,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 508,
+ 186,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 508,
+ 186,
+ 520
+ ],
+ "score": 1.0,
+ "content": "that is not the case.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21.5,
+ "bbox_fs": [
+ 105,
+ 475,
+ 505,
+ 520
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 525,
+ 505,
+ 569
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "The critical insight is that graph filter approximation can be viewed as a polynomial factorization",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 548
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 548
+ ],
+ "score": 1.0,
+ "content": "problem. It is known that any univariate polynomial can be factorized into sub-polynomials of",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 505,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 505,
+ 560
+ ],
+ "score": 1.0,
+ "content": "degree two. Based on this fact, we show by stacking enough SoGCs (and varying their parameters)",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 319,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 319,
+ 570
+ ],
+ "score": 1.0,
+ "content": "can achieve decomposition of any polynomial filters.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 105,
+ 524,
+ 505,
+ 570
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 630
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "In contrast, first-order GCs are not universal approximators; two stacked one-hop GCs cannot model",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "every two-hop GC. Polynomial filter completeness of SoGC leads to better performance of GCNs.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 610
+ ],
+ "score": 1.0,
+ "content": "As shown in Figure 2, networks built with SoGC can overcome over-smoothing and extract features",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 608,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "on high-frequency bands. In the next section, we demonstrate our formal arguments on polynomial",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 619,
+ 169,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 169,
+ 632
+ ],
+ "score": 1.0,
+ "content": "approximation.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 574,
+ 505,
+ 632
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 646,
+ 314,
+ 659
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 315,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 315,
+ 661
+ ],
+ "score": 1.0,
+ "content": "4 REPRESENTATION POWER ANALYSIS",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 671,
+ 298,
+ 683
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 299,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 299,
+ 684
+ ],
+ "score": 1.0,
+ "content": "4.1 LAYER SPANNING SPACE FRAMEWORK",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 692,
+ 504,
+ 715
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 691,
+ 505,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 691,
+ 505,
+ 705
+ ],
+ "score": 1.0,
+ "content": "To illustrate the representation power of GC layers, we establish a Layer Spanning Space (LSS)",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 704,
+ 451,
+ 716
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 704,
+ 451,
+ 716
+ ],
+ "score": 1.0,
+ "content": "framework to study the graph filter space spanned by stacking multiple graph kernels.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 105,
+ 691,
+ 505,
+ 716
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 720,
+ 448,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 450,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 450,
+ 733
+ ],
+ "score": 1.0,
+ "content": "First, we present our mathematical devices in Definition 1, 2 with Lemma 1 as below.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 105,
+ 720,
+ 450,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 117
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 294,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Definition 1. Suppose the parameter space",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 83,
+ 331,
+ 93
+ ],
+ "score": 0.9,
+ "content": "\\Omega \\ = \\ \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 81,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": ". The Linear Shift-Invariant (LSI) graph",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 93,
+ 205,
+ 106
+ ],
+ "score": 1.0,
+ "content": "filter space of degree",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 94,
+ 246,
+ 104
+ ],
+ "score": 0.87,
+ "content": "\\ : \\ : K \\ : \\ : > \\ : \\ : 0 \\ :",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 93,
+ 403,
+ 106
+ ],
+ "score": 1.0,
+ "content": "with respect to a finite graph set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 94,
+ 412,
+ 105
+ ],
+ "score": 0.8,
+ "content": "\\mathcal { G }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 93,
+ 478,
+ 106
+ ],
+ "score": 1.0,
+ "content": "is defined as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 94,
+ 505,
+ 105
+ ],
+ "score": 0.87,
+ "content": "\\begin{array} { r l } { A } & { { } = } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 102,
+ 444,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 104,
+ 254,
+ 118
+ ],
+ "score": 0.85,
+ "content": "\\{ f _ { \\pmb \\theta } : \\hat { \\mathcal { G } } \\times \\mathbb { R } ^ { N } \\mathbb { R } ^ { N } , \\forall \\pmb \\theta \\in \\mathbb { R } ^ { K + 1 } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 102,
+ 433,
+ 118
+ ],
+ "score": 1.0,
+ "content": ", where fθ follows the definition in Equation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 433,
+ 105,
+ 438,
+ 114
+ ],
+ "score": 0.6,
+ "content": "^ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 439,
+ 102,
+ 444,
+ 118
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 118,
+ 505,
+ 152
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 506,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 235,
+ 132
+ ],
+ "score": 1.0,
+ "content": "Definition 2. Let spectrum set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 118,
+ 398,
+ 131
+ ],
+ "score": 0.91,
+ "content": "S ( \\mathcal { G } ) = \\{ \\lambda : \\lambda \\in \\mathcal { S } ( A ( G ) ) , \\forall G \\in \\mathcal { G } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 118,
+ 430,
+ 132
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 118,
+ 455,
+ 130
+ ],
+ "score": 0.91,
+ "content": " { \\mathcal { S } } ( A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 118,
+ 506,
+ 132
+ ],
+ "score": 1.0,
+ "content": "denotes the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 167,
+ 144
+ ],
+ "score": 1.0,
+ "content": "eigenvalues of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 130,
+ 177,
+ 140
+ ],
+ "score": 0.47,
+ "content": "\\pmb { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 128,
+ 288,
+ 144
+ ],
+ "score": 1.0,
+ "content": ". Define spectrum capacity",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 130,
+ 337,
+ 142
+ ],
+ "score": 0.91,
+ "content": "\\Gamma = | S ( \\mathcal { G } ) |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 128,
+ 401,
+ 144
+ ],
+ "score": 1.0,
+ "content": ". In particular,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 130,
+ 470,
+ 142
+ ],
+ "score": 0.91,
+ "content": "\\Gamma = ( N - 1 ) | \\mathcal { G } |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 128,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "if every",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 141,
+ 372,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 372,
+ 154
+ ],
+ "score": 1.0,
+ "content": "graph adjacency matrix has no common eigenvalues other than 1.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 154,
+ 432,
+ 166
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 152,
+ 433,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 152,
+ 203,
+ 168
+ ],
+ "score": 1.0,
+ "content": "Lemma 1. A of degree",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 154,
+ 232,
+ 165
+ ],
+ "score": 0.88,
+ "content": "K > 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 152,
+ 293,
+ 168
+ ],
+ "score": 1.0,
+ "content": "has dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 154,
+ 358,
+ 166
+ ],
+ "score": 0.92,
+ "content": "\\operatorname* { m i n } \\{ K + 1 , \\Gamma \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 152,
+ 433,
+ 168
+ ],
+ "score": 1.0,
+ "content": "as a vector space.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 173,
+ 504,
+ 196
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 505,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 505,
+ 186
+ ],
+ "score": 1.0,
+ "content": "Proof of Lemma 1 follows from Theorem 3 of Sandryhaila & Moura (2013). The complete version",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 184,
+ 468,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 184,
+ 468,
+ 197
+ ],
+ "score": 1.0,
+ "content": "can be found in Appendix C. Here, we induce a finite-dimension filter space by Lemma 1.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 200,
+ 506,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "For simplicity, we will model the linear composition of filters to analyze its representation power.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 212,
+ 504,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 212,
+ 495,
+ 225
+ ],
+ "score": 1.0,
+ "content": "The nonlinear activation effects are beyond the scope of this work. Following Definition 1, let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 213,
+ 504,
+ 222
+ ],
+ "score": 0.79,
+ "content": "\\mathcal { A }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 223,
+ 505,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 235,
+ 235
+ ],
+ "score": 1.0,
+ "content": "be the full filter space of degree",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 223,
+ 261,
+ 234
+ ],
+ "score": 0.87,
+ "content": "\\Gamma - 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 223,
+ 279,
+ 235
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 224,
+ 287,
+ 234
+ ],
+ "score": 0.82,
+ "content": "\\boldsymbol { B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 223,
+ 505,
+ 235
+ ],
+ "score": 1.0,
+ "content": "be the low-level filter space as a set of polynomials in",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 506,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 313,
+ 248
+ ],
+ "score": 1.0,
+ "content": "adjacency matrices (Equation 1). Denote the GC at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 236,
+ 317,
+ 245
+ ],
+ "score": 0.74,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 234,
+ 364,
+ 248
+ ],
+ "score": 1.0,
+ "content": "-th layer by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 234,
+ 400,
+ 247
+ ],
+ "score": 0.92,
+ "content": "f ^ { ( l ) } \\in B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 400,
+ 234,
+ 506,
+ 248
+ ],
+ "score": 1.0,
+ "content": ", then we yield the LSS of",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 246,
+ 218,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 142,
+ 259
+ ],
+ "score": 1.0,
+ "content": "stacking",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 247,
+ 150,
+ 256
+ ],
+ "score": 0.83,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 246,
+ 218,
+ 259
+ ],
+ "score": 1.0,
+ "content": "layers as below:",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 164,
+ 259,
+ 445,
+ 294
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 164,
+ 259,
+ 445,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 164,
+ 259,
+ 445,
+ 294
+ ],
+ "score": 0.95,
+ "content": "\\mathcal { B } ^ { L } = \\left\\{ f : f ( G , \\pmb { x } ) = f ^ { ( L ) } \\circ \\cdots \\circ f ^ { ( 1 ) } ( G , \\pmb { x } ) = \\prod _ { l = 1 } ^ { L } p ^ { ( l ) } ( \\pmb { A } ( G ) ) \\pmb { x } \\right\\} ,",
+ "type": "interline_equation",
+ "image_path": "3df440d997e91fd015e8e28bb9ba44701b6d2620d7f43f17a5e75b42090f22c7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 164,
+ 259,
+ 445,
+ 270.6666666666667
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 164,
+ 270.6666666666667,
+ 445,
+ 282.33333333333337
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 164,
+ 282.33333333333337,
+ 445,
+ 294.00000000000006
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 296,
+ 506,
+ 339
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 506,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 134,
+ 311
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 295,
+ 163,
+ 309
+ ],
+ "score": 0.93,
+ "content": "p ^ { ( l ) } ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 163,
+ 294,
+ 506,
+ 311
+ ],
+ "score": 1.0,
+ "content": "varies in a certain class of polynomials. We can assess the expressive capability of",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 505,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 263,
+ 320
+ ],
+ "score": 1.0,
+ "content": "GC layers by comparing the LSS with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 308,
+ 272,
+ 317
+ ],
+ "score": 0.78,
+ "content": "\\mathcal { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 306,
+ 321,
+ 320
+ ],
+ "score": 1.0,
+ "content": ". Kernels in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 308,
+ 330,
+ 317
+ ],
+ "score": 0.82,
+ "content": "\\boldsymbol { B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 306,
+ 465,
+ 320
+ ],
+ "score": 1.0,
+ "content": "have full representation power if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 307,
+ 501,
+ 318
+ ],
+ "score": 0.9,
+ "content": "{ \\mathcal { A } } \\subseteq { \\mathcal { B } } ^ { L }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 306,
+ 505,
+ 320
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 316,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 188,
+ 331
+ ],
+ "score": 1.0,
+ "content": "We are interested in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 318,
+ 204,
+ 330
+ ],
+ "score": 0.9,
+ "content": "\\boldsymbol { B } _ { K }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 316,
+ 415,
+ 331
+ ],
+ "score": 1.0,
+ "content": ", which denotes all localized filters of degree at most",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 320,
+ 426,
+ 328
+ ],
+ "score": 0.81,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 316,
+ 479,
+ 331
+ ],
+ "score": 1.0,
+ "content": ". The LSS of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 479,
+ 319,
+ 494,
+ 330
+ ],
+ "score": 0.88,
+ "content": "\\boldsymbol { B } _ { K }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 316,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 157,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 157,
+ 342
+ ],
+ "score": 1.0,
+ "content": "modeled as:",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 190,
+ 337,
+ 419,
+ 371
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 190,
+ 337,
+ 419,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 190,
+ 337,
+ 419,
+ 371
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { B } _ { K } ^ { L } = \\left\\{ f : f ( G , \\pmb { x } ) = \\prod _ { l = 1 } ^ { L } \\sum _ { k = 0 } ^ { K } \\theta _ { k } ^ { ( l ) } \\pmb { A } ( G ) ^ { k } \\pmb { x } , \\theta _ { k } ^ { ( l ) } \\in \\mathbb { R } \\right\\} ,",
+ "type": "interline_equation",
+ "image_path": "2a7151fc4b5ede57c94960101aa3e491cdd439cc34dfb2f841d4d1d5b1b2fbb7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 190,
+ 337,
+ 419,
+ 354.0
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 190,
+ 354.0,
+ 419,
+ 371.0
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 371,
+ 450,
+ 384
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 371,
+ 451,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 371,
+ 276,
+ 385
+ ],
+ "score": 1.0,
+ "content": "where the number of layers is bounded by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 372,
+ 351,
+ 384
+ ],
+ "score": 0.9,
+ "content": "L \\le \\lceil ( \\Gamma - 1 ) / K \\rceil",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 352,
+ 371,
+ 451,
+ 385
+ ],
+ "score": 1.0,
+ "content": ", according to Lemma 1.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 395,
+ 339,
+ 407
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 339,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 339,
+ 408
+ ],
+ "score": 1.0,
+ "content": "4.2 UNIVERSAL REPRESENTATION POWER OF SOGC",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 416,
+ 505,
+ 450
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "In this section, we present Theorem 1 to demonstrate the universal representation power of SoGCs",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 428,
+ 504,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 504,
+ 439
+ ],
+ "score": 1.0,
+ "content": "as claimed in Section 3. Formally, we add superscripts to Equation 4 to indicate the layer number.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": "Then we leverage a fundamental polynomial factorization theorem to conclude Theorem 1 as below.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 452,
+ 504,
+ 482
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 450,
+ 504,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 450,
+ 191,
+ 470
+ ],
+ "score": 1.0,
+ "content": "Theorem 1. For any",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 455,
+ 219,
+ 466
+ ],
+ "score": 0.89,
+ "content": "f \\in { \\mathcal { A } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 450,
+ 268,
+ 470
+ ],
+ "score": 1.0,
+ "content": ", there exists",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 451,
+ 308,
+ 467
+ ],
+ "score": 0.92,
+ "content": "f _ { 2 } ^ { ( l ) } \\in B _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 450,
+ 375,
+ 470
+ ],
+ "score": 1.0,
+ "content": "with coefficients",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 451,
+ 504,
+ 467
+ ],
+ "score": 0.92,
+ "content": "\\theta _ { 0 } ^ { ( l ) } , \\theta _ { 1 } ^ { ( l ) } , \\theta _ { 2 } ^ { ( l ) } \\in \\mathbb { R } , l = 1 , \\cdots , L",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 102,
+ 461,
+ 338,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 461,
+ 145,
+ 486
+ ],
+ "score": 1.0,
+ "content": "such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 466,
+ 231,
+ 481
+ ],
+ "score": 0.91,
+ "content": "f = f _ { 2 } ^ { ( L ) } \\circ \\cdots \\circ f _ { 2 } ^ { ( 1 ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 461,
+ 258,
+ 486
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 468,
+ 329,
+ 481
+ ],
+ "score": 0.9,
+ "content": "L \\leq \\lceil ( \\Gamma - 1 ) / 2 \\rceil",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 461,
+ 338,
+ 486
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 487,
+ 504,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 488,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 488,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "The complete proof is presented Appendix D. Theorem 1 can be regarded as the universal approxi-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 499,
+ 496,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 499,
+ 496,
+ 512
+ ],
+ "score": 1.0,
+ "content": "mation theorem of linear GCNs, which implies multi-layer SoGCs have full filter expressiveness.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 515,
+ 505,
+ 550
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 505,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 505,
+ 528
+ ],
+ "score": 1.0,
+ "content": "Theorem 1 also demonstrates how GCNs with SoGCs benefit from depth, which coincides with the",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 506,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 506,
+ 540
+ ],
+ "score": 1.0,
+ "content": "view of Dehmamy et al. (2019). Figure 3a verifies our SoGCN can overcome over-smoothing and",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 319,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 319,
+ 550
+ ],
+ "score": 1.0,
+ "content": "successfully utilize depth to attain performance gain.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 561,
+ 388,
+ 573
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 389,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 389,
+ 574
+ ],
+ "score": 1.0,
+ "content": "4.3 REPRESENTATION POWER OF OTHER GRAPH CONVOLUTION",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 582,
+ 504,
+ 606
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "In this section, we show that vanilla and first-order GCs lack expressiveness, while higher-order GCs",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 594,
+ 308,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 308,
+ 607
+ ],
+ "score": 1.0,
+ "content": "reduce compactness and increase fitting difficulty.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 673
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Vanilla vs. second-order. Extensive works have shown the performance deficiency of vanilla",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 504,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 504,
+ 640
+ ],
+ "score": 1.0,
+ "content": "GCNs (Hoang & Maehara, 2019; Luan et al., 2019; Oono & Suzuki, 2019; Cai & Wang, 2020).",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Based on the LSS framework, we can point out a similar issue but from a novel perspective. Let us",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 648,
+ 507,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 648,
+ 130,
+ 668
+ ],
+ "score": 1.0,
+ "content": "write",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 130,
+ 650,
+ 283,
+ 664
+ ],
+ "score": 0.92,
+ "content": "f _ { 0 } ^ { ( l ) } ( G , \\pmb { x } ) = \\theta ^ { ( l ) } ( \\pmb { A } ( G ) + \\pmb { I } ) \\pmb { x } \\in \\mathcal { B } _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 284,
+ 648,
+ 311,
+ 668
+ ],
+ "score": 1.0,
+ "content": "as the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 653,
+ 316,
+ 662
+ ],
+ "score": 0.73,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 648,
+ 398,
+ 668
+ ],
+ "score": 1.0,
+ "content": "-th GC layer2. Then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 652,
+ 407,
+ 662
+ ],
+ "score": 0.81,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 407,
+ 648,
+ 507,
+ 668
+ ],
+ "score": 1.0,
+ "content": "of them can represent a",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 661,
+ 172,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 661,
+ 172,
+ 675
+ ],
+ "score": 1.0,
+ "content": "LSS as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 212,
+ 671,
+ 398,
+ 705
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 212,
+ 671,
+ 398,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 212,
+ 671,
+ 398,
+ 705
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { B } _ { 0 } ^ { L } = \\left\\{ f : f ( G , \\pmb { x } ) = \\theta \\sum _ { l = 0 } ^ { L } { \\binom { l } { L } \\pmb { A } ( G ) ^ { l } \\pmb { x } } \\right\\} ,",
+ "type": "interline_equation",
+ "image_path": "3f72204e5d3bbfea2fd7c0dd5ff9469275e868bcc4edbaadd88368252a1077a9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 212,
+ 671,
+ 398,
+ 688.0
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 212,
+ 688.0,
+ 398,
+ 705.0
+ ],
+ "spans": [],
+ "index": 44
+ }
+ ]
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 711,
+ 508,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 709,
+ 505,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 709,
+ 193,
+ 724
+ ],
+ "score": 1.0,
+ "content": "2Notice that we use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 712,
+ 205,
+ 721
+ ],
+ "score": 0.86,
+ "content": "\\scriptstyle B _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 205,
+ 709,
+ 505,
+ 724
+ ],
+ "score": 1.0,
+ "content": "to denote the filter space with lumping of self-connection with pairwise neighbor",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 495,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 486,
+ 732
+ ],
+ "score": 1.0,
+ "content": "nodes, since the zero-degree ones are too trivial. We assume the renormalization trick can be merged into",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 487,
+ 722,
+ 492,
+ 730
+ ],
+ "score": 0.79,
+ "content": "\\theta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 492,
+ 721,
+ 495,
+ 732
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "4",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 26,
+ 307,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 117
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 294,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Definition 1. Suppose the parameter space",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 83,
+ 331,
+ 93
+ ],
+ "score": 0.9,
+ "content": "\\Omega \\ = \\ \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 81,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": ". The Linear Shift-Invariant (LSI) graph",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 93,
+ 205,
+ 106
+ ],
+ "score": 1.0,
+ "content": "filter space of degree",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 94,
+ 246,
+ 104
+ ],
+ "score": 0.87,
+ "content": "\\ : \\ : K \\ : \\ : > \\ : \\ : 0 \\ :",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 93,
+ 403,
+ 106
+ ],
+ "score": 1.0,
+ "content": "with respect to a finite graph set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 94,
+ 412,
+ 105
+ ],
+ "score": 0.8,
+ "content": "\\mathcal { G }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 93,
+ 478,
+ 106
+ ],
+ "score": 1.0,
+ "content": "is defined as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 94,
+ 505,
+ 105
+ ],
+ "score": 0.87,
+ "content": "\\begin{array} { r l } { A } & { { } = } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 102,
+ 444,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 104,
+ 254,
+ 118
+ ],
+ "score": 0.85,
+ "content": "\\{ f _ { \\pmb \\theta } : \\hat { \\mathcal { G } } \\times \\mathbb { R } ^ { N } \\mathbb { R } ^ { N } , \\forall \\pmb \\theta \\in \\mathbb { R } ^ { K + 1 } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 102,
+ 433,
+ 118
+ ],
+ "score": 1.0,
+ "content": ", where fθ follows the definition in Equation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 433,
+ 105,
+ 438,
+ 114
+ ],
+ "score": 0.6,
+ "content": "^ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 439,
+ 102,
+ 444,
+ 118
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 104,
+ 81,
+ 506,
+ 118
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 118,
+ 505,
+ 152
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 506,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 118,
+ 235,
+ 132
+ ],
+ "score": 1.0,
+ "content": "Definition 2. Let spectrum set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 118,
+ 398,
+ 131
+ ],
+ "score": 0.91,
+ "content": "S ( \\mathcal { G } ) = \\{ \\lambda : \\lambda \\in \\mathcal { S } ( A ( G ) ) , \\forall G \\in \\mathcal { G } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 118,
+ 430,
+ 132
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 118,
+ 455,
+ 130
+ ],
+ "score": 0.91,
+ "content": " { \\mathcal { S } } ( A )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 118,
+ 506,
+ 132
+ ],
+ "score": 1.0,
+ "content": "denotes the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 167,
+ 144
+ ],
+ "score": 1.0,
+ "content": "eigenvalues of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 130,
+ 177,
+ 140
+ ],
+ "score": 0.47,
+ "content": "\\pmb { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 128,
+ 288,
+ 144
+ ],
+ "score": 1.0,
+ "content": ". Define spectrum capacity",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 130,
+ 337,
+ 142
+ ],
+ "score": 0.91,
+ "content": "\\Gamma = | S ( \\mathcal { G } ) |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 128,
+ 401,
+ 144
+ ],
+ "score": 1.0,
+ "content": ". In particular,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 130,
+ 470,
+ 142
+ ],
+ "score": 0.91,
+ "content": "\\Gamma = ( N - 1 ) | \\mathcal { G } |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 128,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "if every",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 141,
+ 372,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 372,
+ 154
+ ],
+ "score": 1.0,
+ "content": "graph adjacency matrix has no common eigenvalues other than 1.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 105,
+ 118,
+ 506,
+ 154
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 154,
+ 432,
+ 166
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 152,
+ 433,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 152,
+ 203,
+ 168
+ ],
+ "score": 1.0,
+ "content": "Lemma 1. A of degree",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 154,
+ 232,
+ 165
+ ],
+ "score": 0.88,
+ "content": "K > 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 152,
+ 293,
+ 168
+ ],
+ "score": 1.0,
+ "content": "has dimension",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 294,
+ 154,
+ 358,
+ 166
+ ],
+ "score": 0.92,
+ "content": "\\operatorname* { m i n } \\{ K + 1 , \\Gamma \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 152,
+ 433,
+ 168
+ ],
+ "score": 1.0,
+ "content": "as a vector space.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 152,
+ 433,
+ 168
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 173,
+ 504,
+ 196
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 505,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 505,
+ 186
+ ],
+ "score": 1.0,
+ "content": "Proof of Lemma 1 follows from Theorem 3 of Sandryhaila & Moura (2013). The complete version",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 184,
+ 468,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 184,
+ 468,
+ 197
+ ],
+ "score": 1.0,
+ "content": "can be found in Appendix C. Here, we induce a finite-dimension filter space by Lemma 1.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 105,
+ 173,
+ 505,
+ 197
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 200,
+ 506,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "For simplicity, we will model the linear composition of filters to analyze its representation power.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 212,
+ 504,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 212,
+ 495,
+ 225
+ ],
+ "score": 1.0,
+ "content": "The nonlinear activation effects are beyond the scope of this work. Following Definition 1, let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 213,
+ 504,
+ 222
+ ],
+ "score": 0.79,
+ "content": "\\mathcal { A }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 223,
+ 505,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 235,
+ 235
+ ],
+ "score": 1.0,
+ "content": "be the full filter space of degree",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 223,
+ 261,
+ 234
+ ],
+ "score": 0.87,
+ "content": "\\Gamma - 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 223,
+ 279,
+ 235
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 224,
+ 287,
+ 234
+ ],
+ "score": 0.82,
+ "content": "\\boldsymbol { B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 223,
+ 505,
+ 235
+ ],
+ "score": 1.0,
+ "content": "be the low-level filter space as a set of polynomials in",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 506,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 313,
+ 248
+ ],
+ "score": 1.0,
+ "content": "adjacency matrices (Equation 1). Denote the GC at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 236,
+ 317,
+ 245
+ ],
+ "score": 0.74,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 234,
+ 364,
+ 248
+ ],
+ "score": 1.0,
+ "content": "-th layer by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 365,
+ 234,
+ 400,
+ 247
+ ],
+ "score": 0.92,
+ "content": "f ^ { ( l ) } \\in B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 400,
+ 234,
+ 506,
+ 248
+ ],
+ "score": 1.0,
+ "content": ", then we yield the LSS of",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 246,
+ 218,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 142,
+ 259
+ ],
+ "score": 1.0,
+ "content": "stacking",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 247,
+ 150,
+ 256
+ ],
+ "score": 0.83,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 246,
+ 218,
+ 259
+ ],
+ "score": 1.0,
+ "content": "layers as below:",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 201,
+ 506,
+ 259
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 164,
+ 259,
+ 445,
+ 294
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 164,
+ 259,
+ 445,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 164,
+ 259,
+ 445,
+ 294
+ ],
+ "score": 0.95,
+ "content": "\\mathcal { B } ^ { L } = \\left\\{ f : f ( G , \\pmb { x } ) = f ^ { ( L ) } \\circ \\cdots \\circ f ^ { ( 1 ) } ( G , \\pmb { x } ) = \\prod _ { l = 1 } ^ { L } p ^ { ( l ) } ( \\pmb { A } ( G ) ) \\pmb { x } \\right\\} ,",
+ "type": "interline_equation",
+ "image_path": "3df440d997e91fd015e8e28bb9ba44701b6d2620d7f43f17a5e75b42090f22c7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 164,
+ 259,
+ 445,
+ 270.6666666666667
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 164,
+ 270.6666666666667,
+ 445,
+ 282.33333333333337
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 164,
+ 282.33333333333337,
+ 445,
+ 294.00000000000006
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 296,
+ 506,
+ 339
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 506,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 294,
+ 134,
+ 311
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 295,
+ 163,
+ 309
+ ],
+ "score": 0.93,
+ "content": "p ^ { ( l ) } ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 163,
+ 294,
+ 506,
+ 311
+ ],
+ "score": 1.0,
+ "content": "varies in a certain class of polynomials. We can assess the expressive capability of",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 505,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 263,
+ 320
+ ],
+ "score": 1.0,
+ "content": "GC layers by comparing the LSS with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 308,
+ 272,
+ 317
+ ],
+ "score": 0.78,
+ "content": "\\mathcal { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 306,
+ 321,
+ 320
+ ],
+ "score": 1.0,
+ "content": ". Kernels in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 322,
+ 308,
+ 330,
+ 317
+ ],
+ "score": 0.82,
+ "content": "\\boldsymbol { B }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 306,
+ 465,
+ 320
+ ],
+ "score": 1.0,
+ "content": "have full representation power if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 307,
+ 501,
+ 318
+ ],
+ "score": 0.9,
+ "content": "{ \\mathcal { A } } \\subseteq { \\mathcal { B } } ^ { L }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 306,
+ 505,
+ 320
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 316,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 188,
+ 331
+ ],
+ "score": 1.0,
+ "content": "We are interested in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 318,
+ 204,
+ 330
+ ],
+ "score": 0.9,
+ "content": "\\boldsymbol { B } _ { K }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 204,
+ 316,
+ 415,
+ 331
+ ],
+ "score": 1.0,
+ "content": ", which denotes all localized filters of degree at most",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 320,
+ 426,
+ 328
+ ],
+ "score": 0.81,
+ "content": "K",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 426,
+ 316,
+ 479,
+ 331
+ ],
+ "score": 1.0,
+ "content": ". The LSS of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 479,
+ 319,
+ 494,
+ 330
+ ],
+ "score": 0.88,
+ "content": "\\boldsymbol { B } _ { K }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 316,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 157,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 157,
+ 342
+ ],
+ "score": 1.0,
+ "content": "modeled as:",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 105,
+ 294,
+ 506,
+ 342
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 190,
+ 337,
+ 419,
+ 371
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 190,
+ 337,
+ 419,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 190,
+ 337,
+ 419,
+ 371
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { B } _ { K } ^ { L } = \\left\\{ f : f ( G , \\pmb { x } ) = \\prod _ { l = 1 } ^ { L } \\sum _ { k = 0 } ^ { K } \\theta _ { k } ^ { ( l ) } \\pmb { A } ( G ) ^ { k } \\pmb { x } , \\theta _ { k } ^ { ( l ) } \\in \\mathbb { R } \\right\\} ,",
+ "type": "interline_equation",
+ "image_path": "2a7151fc4b5ede57c94960101aa3e491cdd439cc34dfb2f841d4d1d5b1b2fbb7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 190,
+ 337,
+ 419,
+ 354.0
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 190,
+ 354.0,
+ 419,
+ 371.0
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 371,
+ 450,
+ 384
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 371,
+ 451,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 371,
+ 276,
+ 385
+ ],
+ "score": 1.0,
+ "content": "where the number of layers is bounded by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 372,
+ 351,
+ 384
+ ],
+ "score": 0.9,
+ "content": "L \\le \\lceil ( \\Gamma - 1 ) / K \\rceil",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 352,
+ 371,
+ 451,
+ 385
+ ],
+ "score": 1.0,
+ "content": ", according to Lemma 1.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 106,
+ 371,
+ 451,
+ 385
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 395,
+ 339,
+ 407
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 339,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 339,
+ 408
+ ],
+ "score": 1.0,
+ "content": "4.2 UNIVERSAL REPRESENTATION POWER OF SOGC",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 416,
+ 505,
+ 450
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "In this section, we present Theorem 1 to demonstrate the universal representation power of SoGCs",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 428,
+ 504,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 504,
+ 439
+ ],
+ "score": 1.0,
+ "content": "as claimed in Section 3. Formally, we add superscripts to Equation 4 to indicate the layer number.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": "Then we leverage a fundamental polynomial factorization theorem to conclude Theorem 1 as below.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 106,
+ 417,
+ 505,
+ 451
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 452,
+ 504,
+ 482
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 450,
+ 504,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 450,
+ 191,
+ 470
+ ],
+ "score": 1.0,
+ "content": "Theorem 1. For any",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 455,
+ 219,
+ 466
+ ],
+ "score": 0.89,
+ "content": "f \\in { \\mathcal { A } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 450,
+ 268,
+ 470
+ ],
+ "score": 1.0,
+ "content": ", there exists",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 451,
+ 308,
+ 467
+ ],
+ "score": 0.92,
+ "content": "f _ { 2 } ^ { ( l ) } \\in B _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 450,
+ 375,
+ 470
+ ],
+ "score": 1.0,
+ "content": "with coefficients",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 451,
+ 504,
+ 467
+ ],
+ "score": 0.92,
+ "content": "\\theta _ { 0 } ^ { ( l ) } , \\theta _ { 1 } ^ { ( l ) } , \\theta _ { 2 } ^ { ( l ) } \\in \\mathbb { R } , l = 1 , \\cdots , L",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 102,
+ 461,
+ 338,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 102,
+ 461,
+ 145,
+ 486
+ ],
+ "score": 1.0,
+ "content": "such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 466,
+ 231,
+ 481
+ ],
+ "score": 0.91,
+ "content": "f = f _ { 2 } ^ { ( L ) } \\circ \\cdots \\circ f _ { 2 } ^ { ( 1 ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 231,
+ 461,
+ 258,
+ 486
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 468,
+ 329,
+ 481
+ ],
+ "score": 0.9,
+ "content": "L \\leq \\lceil ( \\Gamma - 1 ) / 2 \\rceil",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 461,
+ 338,
+ 486
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 102,
+ 450,
+ 504,
+ 486
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 487,
+ 504,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 488,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 488,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "The complete proof is presented Appendix D. Theorem 1 can be regarded as the universal approxi-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 499,
+ 496,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 499,
+ 496,
+ 512
+ ],
+ "score": 1.0,
+ "content": "mation theorem of linear GCNs, which implies multi-layer SoGCs have full filter expressiveness.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 106,
+ 488,
+ 505,
+ 512
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 515,
+ 505,
+ 550
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 505,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 505,
+ 528
+ ],
+ "score": 1.0,
+ "content": "Theorem 1 also demonstrates how GCNs with SoGCs benefit from depth, which coincides with the",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 506,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 506,
+ 540
+ ],
+ "score": 1.0,
+ "content": "view of Dehmamy et al. (2019). Figure 3a verifies our SoGCN can overcome over-smoothing and",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 319,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 319,
+ 550
+ ],
+ "score": 1.0,
+ "content": "successfully utilize depth to attain performance gain.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 105,
+ 515,
+ 506,
+ 550
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 561,
+ 388,
+ 573
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 389,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 389,
+ 574
+ ],
+ "score": 1.0,
+ "content": "4.3 REPRESENTATION POWER OF OTHER GRAPH CONVOLUTION",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 582,
+ 504,
+ 606
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 582,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "In this section, we show that vanilla and first-order GCs lack expressiveness, while higher-order GCs",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 594,
+ 308,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 308,
+ 607
+ ],
+ "score": 1.0,
+ "content": "reduce compactness and increase fitting difficulty.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 106,
+ 582,
+ 505,
+ 607
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 673
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Vanilla vs. second-order. Extensive works have shown the performance deficiency of vanilla",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 504,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 504,
+ 640
+ ],
+ "score": 1.0,
+ "content": "GCNs (Hoang & Maehara, 2019; Luan et al., 2019; Oono & Suzuki, 2019; Cai & Wang, 2020).",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Based on the LSS framework, we can point out a similar issue but from a novel perspective. Let us",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 648,
+ 507,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 648,
+ 130,
+ 668
+ ],
+ "score": 1.0,
+ "content": "write",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 130,
+ 650,
+ 283,
+ 664
+ ],
+ "score": 0.92,
+ "content": "f _ { 0 } ^ { ( l ) } ( G , \\pmb { x } ) = \\theta ^ { ( l ) } ( \\pmb { A } ( G ) + \\pmb { I } ) \\pmb { x } \\in \\mathcal { B } _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 284,
+ 648,
+ 311,
+ 668
+ ],
+ "score": 1.0,
+ "content": "as the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 653,
+ 316,
+ 662
+ ],
+ "score": 0.73,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 648,
+ 398,
+ 668
+ ],
+ "score": 1.0,
+ "content": "-th GC layer2. Then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 399,
+ 652,
+ 407,
+ 662
+ ],
+ "score": 0.81,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 407,
+ 648,
+ 507,
+ 668
+ ],
+ "score": 1.0,
+ "content": "of them can represent a",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 661,
+ 172,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 661,
+ 172,
+ 675
+ ],
+ "score": 1.0,
+ "content": "LSS as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 105,
+ 615,
+ 507,
+ 675
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 212,
+ 671,
+ 398,
+ 705
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 212,
+ 671,
+ 398,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 212,
+ 671,
+ 398,
+ 705
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { B } _ { 0 } ^ { L } = \\left\\{ f : f ( G , \\pmb { x } ) = \\theta \\sum _ { l = 0 } ^ { L } { \\binom { l } { L } \\pmb { A } ( G ) ^ { l } \\pmb { x } } \\right\\} ,",
+ "type": "interline_equation",
+ "image_path": "3f72204e5d3bbfea2fd7c0dd5ff9469275e868bcc4edbaadd88368252a1077a9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 43.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 212,
+ 671,
+ 398,
+ 688.0
+ ],
+ "spans": [],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 212,
+ 688.0,
+ 398,
+ 705.0
+ ],
+ "spans": [],
+ "index": 44
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 147,
+ 95
+ ],
+ "score": 1.0,
+ "content": "by letting",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 147,
+ 81,
+ 214,
+ 92
+ ],
+ "score": 0.91,
+ "content": "\\theta = \\theta ^ { ( L ) } \\cdot \\cdot \\cdot \\theta ^ { ( 1 ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 215,
+ 80,
+ 304,
+ 95
+ ],
+ "score": 1.0,
+ "content": ". No matter how large",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 83,
+ 312,
+ 92
+ ],
+ "score": 0.8,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 80,
+ 486,
+ 95
+ ],
+ "score": 1.0,
+ "content": "is or how a optimizer tunes the parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 486,
+ 81,
+ 501,
+ 92
+ ],
+ "score": 0.88,
+ "content": "\\theta ^ { ( l ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 80,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 92,
+ 408,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 93,
+ 157,
+ 105
+ ],
+ "score": 0.92,
+ "content": "\\dot { \\dim } { \\cal B } _ { 0 } ^ { L } \\dot { = } 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 92,
+ 221,
+ 107
+ ],
+ "score": 1.0,
+ "content": "which signifies",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 221,
+ 93,
+ 235,
+ 105
+ ],
+ "score": 0.9,
+ "content": "B _ { 0 } ^ { L }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 92,
+ 394,
+ 107
+ ],
+ "score": 1.0,
+ "content": "degenerates to a negligible subspace of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 394,
+ 94,
+ 403,
+ 104
+ ],
+ "score": 0.83,
+ "content": "\\mathcal { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 403,
+ 92,
+ 408,
+ 107
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 116,
+ 503,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 504,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 354,
+ 133
+ ],
+ "score": 1.0,
+ "content": "First-order vs. second-order. We denote first-order GCs as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 355,
+ 116,
+ 504,
+ 132
+ ],
+ "score": 0.91,
+ "content": "f _ { 1 } ^ { ( l ) } ( G , { \\pmb x } ) = ( \\theta _ { 1 } ^ { ( l ) } { \\pmb A } ( G ) + \\theta _ { 0 } ^ { ( l ) } { \\pmb I } ) { \\pmb x } \\in",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 129,
+ 303,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 131,
+ 118,
+ 141
+ ],
+ "score": 0.87,
+ "content": "\\boldsymbol { B } _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 129,
+ 303,
+ 142
+ ],
+ "score": 1.0,
+ "content": ". In the spirit of Section 4.1, write the LSS as:",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 169,
+ 143,
+ 440,
+ 178
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 169,
+ 143,
+ 440,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 169,
+ 143,
+ 440,
+ 178
+ ],
+ "score": 0.94,
+ "content": "\\mathcal { B } _ { 1 } ^ { L } = \\left\\{ f : f ( G , \\pmb { x } ) = \\prod _ { l = 1 } ^ { L } \\left( \\theta _ { 1 } ^ { ( l ) } A ( G ) + \\theta _ { 0 } ^ { ( l ) } \\pmb { I } \\right) \\pmb { x } , \\theta _ { 0 } ^ { ( l ) } , \\theta _ { 1 } ^ { ( l ) } \\in \\mathbb { R } \\right\\} ,",
+ "type": "interline_equation",
+ "image_path": "fb5047d3b964cf275ec14c3f0dadbca146fbac84a0e270919d451d517f905f9b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 169,
+ 143,
+ 440,
+ 154.66666666666666
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 169,
+ 154.66666666666666,
+ 440,
+ 166.33333333333331
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 169,
+ 166.33333333333331,
+ 440,
+ 177.99999999999997
+ ],
+ "spans": [],
+ "index": 6
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 178,
+ 505,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 178,
+ 506,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 178,
+ 506,
+ 191
+ ],
+ "score": 1.0,
+ "content": "which is isomorphic to a polynomial space whose elements split over the real domain. Compared",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 506,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 127,
+ 203
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 189,
+ 141,
+ 201
+ ],
+ "score": 0.9,
+ "content": "B _ { 0 } ^ { L }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 141,
+ 188,
+ 198,
+ 203
+ ],
+ "score": 1.0,
+ "content": "(Equation 7),",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 189,
+ 212,
+ 201
+ ],
+ "score": 0.9,
+ "content": "\\vec { B _ { 1 } ^ { L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 188,
+ 354,
+ 203
+ ],
+ "score": 1.0,
+ "content": "represents a much larger subset of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 190,
+ 363,
+ 199
+ ],
+ "score": 0.81,
+ "content": "\\mathcal { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 363,
+ 188,
+ 506,
+ 203
+ ],
+ "score": 1.0,
+ "content": ". This highlights the importance of",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 506,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 343,
+ 213
+ ],
+ "score": 1.0,
+ "content": "the first-order term or the identity mapping mentioned in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 201,
+ 357,
+ 210
+ ],
+ "score": 0.43,
+ "content": "\\mathrm { X u }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 199,
+ 506,
+ 213
+ ],
+ "score": 1.0,
+ "content": "et al., 2019; Dehmamy et al., 2019;",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 211,
+ 207,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 211,
+ 207,
+ 223
+ ],
+ "score": 1.0,
+ "content": "Ming Chen et al., 2020).",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 228,
+ 505,
+ 262
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 228,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 228,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "The limitations also become obvious since not all polynomials can be factorized into first-order",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 239,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 239,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "polynomials. These polynomials only occupy a small proportion in the ambient polynomial space",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 250,
+ 452,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 452,
+ 263
+ ],
+ "score": 1.0,
+ "content": "(Li, 2011), which indicates first-order GCs are not universal approximators in general.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 272,
+ 505,
+ 394
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 318,
+ 286
+ ],
+ "score": 1.0,
+ "content": "Higher-order vs. second-order. GCs of degree",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 273,
+ 354,
+ 284
+ ],
+ "score": 0.91,
+ "content": "K \\ \\geq \\ 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 354,
+ 272,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "are called higher-order GCs. They",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 284,
+ 505,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 505,
+ 297
+ ],
+ "score": 1.0,
+ "content": "can model multi-hop GCNs such as Luan et al. (2019); Liao et al. (2019); Abu-El-Haija et al.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 308
+ ],
+ "score": 1.0,
+ "content": "(2019). Higher-order GCs have equivalent expressive power to SoGCs, since they can be reduced to",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 320
+ ],
+ "score": 1.0,
+ "content": "SoGCs as long as coefficient sparsity can be achieved. But this by-product–an uncertain sparsity of",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 317,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 317,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "coefficients–is not compatible with gradient-based optimization algorithms. Extensive experiments",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 327,
+ 506,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 327,
+ 506,
+ 340
+ ],
+ "score": 1.0,
+ "content": "(Defferrard et al., 2016) have shown the ineffectiveness of learning higher-order kernels, because",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 338,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "eigenvalues of graph adjacency matrices diminish when powered. This results in a decreasing nu-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 169,
+ 362
+ ],
+ "score": 1.0,
+ "content": "merical rank of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 170,
+ 349,
+ 199,
+ 362
+ ],
+ "score": 0.92,
+ "content": "{ \\bar { \\mathbf { A } } } ( G ) ^ { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 349,
+ 505,
+ 362
+ ],
+ "score": 1.0,
+ "content": ", which prevent higher-order GCs from aggregating larger-scale information.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 359,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "SoGCs can alleviate this problem by preventing the loss of information due to higher-order powering",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 384
+ ],
+ "score": 1.0,
+ "content": "operation. Finally, higher-order GC lacks nonlinearity. SoGCN can bring a better balance between",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 390,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 390,
+ 396
+ ],
+ "score": 1.0,
+ "content": "the expressive power of low-level layers and nonlinearity among them.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 409,
+ 408,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 409,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 409,
+ 424
+ ],
+ "score": 1.0,
+ "content": "5 SECOND-ORDER GRAPH CONVOLUTIONAL NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 433,
+ 505,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 504,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 504,
+ 445
+ ],
+ "score": 1.0,
+ "content": "In this section, we introduce other building blocks of GCNs and establish our Second-Order Graph",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Convolutional Networks (SoGCN) following the fashion of deep learning. First, we promote SoGC",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": "to the multi-channel version analogous to Kipf & Welling (2017). Then we cascade a feature embed-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "ding layer, multiple SoGC layers, and append a readout module. Suppose the multi-channel input is",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 107,
+ 475,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 478,
+ 158,
+ 489
+ ],
+ "score": 0.91,
+ "content": "\\pmb { X } \\in \\mathbb { R } ^ { N \\times D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 159,
+ 475,
+ 238,
+ 492
+ ],
+ "score": 1.0,
+ "content": "supported in graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 478,
+ 267,
+ 489
+ ],
+ "score": 0.9,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 267,
+ 475,
+ 354,
+ 492
+ ],
+ "score": 1.0,
+ "content": ", denote the output of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 479,
+ 359,
+ 488
+ ],
+ "score": 0.68,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 359,
+ 475,
+ 405,
+ 492
+ ],
+ "score": 1.0,
+ "content": "-th layer as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 477,
+ 466,
+ 489
+ ],
+ "score": 0.92,
+ "content": "\\pmb { X } ^ { ( l ) } \\in \\mathbb { R } ^ { N \\times E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 475,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": ", the final",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 103,
+ 485,
+ 507,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 485,
+ 191,
+ 504
+ ],
+ "score": 1.0,
+ "content": "node-level output as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 191,
+ 489,
+ 241,
+ 500
+ ],
+ "score": 0.92,
+ "content": "\\pmb { Y } \\in \\bar { \\mathbb { R } } ^ { N \\times F }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 242,
+ 485,
+ 345,
+ 504
+ ],
+ "score": 1.0,
+ "content": ", or graph-level output as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 488,
+ 382,
+ 500
+ ],
+ "score": 0.91,
+ "content": "\\pmb { Y } \\in \\mathbb { R } ^ { \\pmb { E } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 485,
+ 507,
+ 504
+ ],
+ "score": 1.0,
+ "content": ", we formulate our novel deep",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 241,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 241,
+ 513
+ ],
+ "score": 1.0,
+ "content": "GCN based on SoGC as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 159,
+ 511,
+ 450,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 159,
+ 511,
+ 450,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 159,
+ 511,
+ 450,
+ 572
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { \\pmb { X } ^ { ( 0 ) } = \\rho \\left( \\pmb { X } ; \\pmb { \\Phi } \\right) , } \\\\ & { \\pmb { X } ^ { ( l + 1 ) } = \\sigma \\left( \\pmb { A } ( G ) ^ { 2 } \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 2 } ^ { ( l + 1 ) } + \\pmb { A } ( G ) \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 1 } ^ { ( l + 1 ) } + \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 0 } ^ { ( l + 1 ) } \\right) , } \\\\ & { \\pmb { Y } = \\tau \\left( \\pmb { X } ^ { ( L ) } ; \\pmb { \\Psi } \\right) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "9780b1edac100fa3f4a86bb987250b3f51c59b8f1a3ff4f05c9d9a0912c717d9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 34,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 159,
+ 511,
+ 450,
+ 531.3333333333334
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 159,
+ 531.3333333333334,
+ 450,
+ 551.6666666666667
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 159,
+ 551.6666666666667,
+ 450,
+ 572.0000000000001
+ ],
+ "spans": [],
+ "index": 35
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 572,
+ 504,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 572,
+ 135,
+ 590
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 573,
+ 246,
+ 588
+ ],
+ "score": 0.92,
+ "content": "\\Theta _ { i } ^ { ( l ) } \\in \\mathbb { R } ^ { E \\times E } , i = 0 , 1 , 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 572,
+ 409,
+ 590
+ ],
+ "score": 1.0,
+ "content": "are trainable weights for linear filters;",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 410,
+ 575,
+ 504,
+ 587
+ ],
+ "score": 0.89,
+ "content": "\\rho : \\mathbb { R } ^ { N \\times D } \\mathbb { R } ^ { N \\times E }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 586,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 586,
+ 484,
+ 601
+ ],
+ "score": 1.0,
+ "content": "is an equivariant embedder (Maron et al., 2018) with parameters Φ; σ : RN×E → RN×E",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 481,
+ 587,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "is an",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 596,
+ 507,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 596,
+ 281,
+ 613
+ ],
+ "score": 1.0,
+ "content": "activation function. For node-level readout,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 598,
+ 366,
+ 609
+ ],
+ "score": 0.91,
+ "content": "\\tau : \\mathbb { R } ^ { N \\times E } \\stackrel { \\bullet } { \\to } \\mathbb { R } ^ { N \\times F }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 366,
+ 596,
+ 507,
+ 613
+ ],
+ "score": 1.0,
+ "content": "can be a decoder (with parameters",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 608,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 610,
+ 117,
+ 620
+ ],
+ "score": 0.64,
+ "content": "\\Psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 117,
+ 608,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": ") or a nonlinear activation (e.g., softmax) in place of the prior layer. For graph-level output,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 107,
+ 617,
+ 507,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 620,
+ 181,
+ 631
+ ],
+ "score": 0.89,
+ "content": "\\tau : \\mathbb { R } ^ { N \\times E } \\mathbb { R } ^ { E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 617,
+ 507,
+ 635
+ ],
+ "score": 1.0,
+ "content": "should be an invariant readout function (Maron et al., 2018), e.g., channel-wise",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 631,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 631,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "sum, mean or max (Hamilton et al., 2017). In practice, we adopt ReLU as nonlinear activation (i.e.,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 107,
+ 643,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 643,
+ 155,
+ 654
+ ],
+ "score": 0.82,
+ "content": "\\sigma = \\mathrm { R e L U }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 643,
+ 401,
+ 655
+ ],
+ "score": 1.0,
+ "content": "), a multi-layer perceptron (MLP) as the embedding function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 402,
+ 645,
+ 408,
+ 654
+ ],
+ "score": 0.76,
+ "content": "\\rho",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 408,
+ 643,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": ", another MLP for node",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 417,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 224,
+ 666
+ ],
+ "score": 1.0,
+ "content": "regression readout, and sum (",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 224,
+ 654,
+ 238,
+ 664
+ ],
+ "score": 0.38,
+ "content": "\\mathrm { { X u } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 654,
+ 417,
+ 666
+ ],
+ "score": 1.0,
+ "content": "et al., 2019) for graph classification readout.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 39.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 678,
+ 243,
+ 689
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 678,
+ 244,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 678,
+ 244,
+ 691
+ ],
+ "score": 1.0,
+ "content": "5.1 GATED RECURRENT UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 698,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "Gated Recurrent Unit (GRU) has been served as a basic building block in message-passing GNN",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "architectures (Li et al., 2016; Gilmer et al., 2017; Corso et al., 2020). In this subsection, we explore",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 238,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 238,
+ 733
+ ],
+ "score": 1.0,
+ "content": "its application in spectral GCNs.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 26,
+ 307,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 147,
+ 95
+ ],
+ "score": 1.0,
+ "content": "by letting",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 147,
+ 81,
+ 214,
+ 92
+ ],
+ "score": 0.91,
+ "content": "\\theta = \\theta ^ { ( L ) } \\cdot \\cdot \\cdot \\theta ^ { ( 1 ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 215,
+ 80,
+ 304,
+ 95
+ ],
+ "score": 1.0,
+ "content": ". No matter how large",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 83,
+ 312,
+ 92
+ ],
+ "score": 0.8,
+ "content": "L",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 80,
+ 486,
+ 95
+ ],
+ "score": 1.0,
+ "content": "is or how a optimizer tunes the parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 486,
+ 81,
+ 501,
+ 92
+ ],
+ "score": 0.88,
+ "content": "\\theta ^ { ( l ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 80,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 92,
+ 408,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 93,
+ 157,
+ 105
+ ],
+ "score": 0.92,
+ "content": "\\dot { \\dim } { \\cal B } _ { 0 } ^ { L } \\dot { = } 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 92,
+ 221,
+ 107
+ ],
+ "score": 1.0,
+ "content": "which signifies",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 221,
+ 93,
+ 235,
+ 105
+ ],
+ "score": 0.9,
+ "content": "B _ { 0 } ^ { L }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 92,
+ 394,
+ 107
+ ],
+ "score": 1.0,
+ "content": "degenerates to a negligible subspace of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 394,
+ 94,
+ 403,
+ 104
+ ],
+ "score": 0.83,
+ "content": "\\mathcal { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 403,
+ 92,
+ 408,
+ 107
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 80,
+ 505,
+ 107
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 116,
+ 503,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 504,
+ 133
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 354,
+ 133
+ ],
+ "score": 1.0,
+ "content": "First-order vs. second-order. We denote first-order GCs as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 355,
+ 116,
+ 504,
+ 132
+ ],
+ "score": 0.91,
+ "content": "f _ { 1 } ^ { ( l ) } ( G , { \\pmb x } ) = ( \\theta _ { 1 } ^ { ( l ) } { \\pmb A } ( G ) + \\theta _ { 0 } ^ { ( l ) } { \\pmb I } ) { \\pmb x } \\in",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 129,
+ 303,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 131,
+ 118,
+ 141
+ ],
+ "score": 0.87,
+ "content": "\\boldsymbol { B } _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 129,
+ 303,
+ 142
+ ],
+ "score": 1.0,
+ "content": ". In the spirit of Section 4.1, write the LSS as:",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 115,
+ 504,
+ 142
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 169,
+ 143,
+ 440,
+ 178
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 169,
+ 143,
+ 440,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 169,
+ 143,
+ 440,
+ 178
+ ],
+ "score": 0.94,
+ "content": "\\mathcal { B } _ { 1 } ^ { L } = \\left\\{ f : f ( G , \\pmb { x } ) = \\prod _ { l = 1 } ^ { L } \\left( \\theta _ { 1 } ^ { ( l ) } A ( G ) + \\theta _ { 0 } ^ { ( l ) } \\pmb { I } \\right) \\pmb { x } , \\theta _ { 0 } ^ { ( l ) } , \\theta _ { 1 } ^ { ( l ) } \\in \\mathbb { R } \\right\\} ,",
+ "type": "interline_equation",
+ "image_path": "fb5047d3b964cf275ec14c3f0dadbca146fbac84a0e270919d451d517f905f9b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 169,
+ 143,
+ 440,
+ 154.66666666666666
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 169,
+ 154.66666666666666,
+ 440,
+ 166.33333333333331
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 169,
+ 166.33333333333331,
+ 440,
+ 177.99999999999997
+ ],
+ "spans": [],
+ "index": 6
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 178,
+ 505,
+ 223
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 178,
+ 506,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 178,
+ 506,
+ 191
+ ],
+ "score": 1.0,
+ "content": "which is isomorphic to a polynomial space whose elements split over the real domain. Compared",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 506,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 127,
+ 203
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 189,
+ 141,
+ 201
+ ],
+ "score": 0.9,
+ "content": "B _ { 0 } ^ { L }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 141,
+ 188,
+ 198,
+ 203
+ ],
+ "score": 1.0,
+ "content": "(Equation 7),",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 189,
+ 212,
+ 201
+ ],
+ "score": 0.9,
+ "content": "\\vec { B _ { 1 } ^ { L } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 188,
+ 354,
+ 203
+ ],
+ "score": 1.0,
+ "content": "represents a much larger subset of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 190,
+ 363,
+ 199
+ ],
+ "score": 0.81,
+ "content": "\\mathcal { A }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 363,
+ 188,
+ 506,
+ 203
+ ],
+ "score": 1.0,
+ "content": ". This highlights the importance of",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 506,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 343,
+ 213
+ ],
+ "score": 1.0,
+ "content": "the first-order term or the identity mapping mentioned in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 201,
+ 357,
+ 210
+ ],
+ "score": 0.43,
+ "content": "\\mathrm { X u }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 199,
+ 506,
+ 213
+ ],
+ "score": 1.0,
+ "content": "et al., 2019; Dehmamy et al., 2019;",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 211,
+ 207,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 211,
+ 207,
+ 223
+ ],
+ "score": 1.0,
+ "content": "Ming Chen et al., 2020).",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 105,
+ 178,
+ 506,
+ 223
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 228,
+ 505,
+ 262
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 228,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 228,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "The limitations also become obvious since not all polynomials can be factorized into first-order",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 239,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 239,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "polynomials. These polynomials only occupy a small proportion in the ambient polynomial space",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 250,
+ 452,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 452,
+ 263
+ ],
+ "score": 1.0,
+ "content": "(Li, 2011), which indicates first-order GCs are not universal approximators in general.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 106,
+ 228,
+ 505,
+ 263
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 272,
+ 505,
+ 394
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 318,
+ 286
+ ],
+ "score": 1.0,
+ "content": "Higher-order vs. second-order. GCs of degree",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 318,
+ 273,
+ 354,
+ 284
+ ],
+ "score": 0.91,
+ "content": "K \\ \\geq \\ 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 354,
+ 272,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "are called higher-order GCs. They",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 284,
+ 505,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 284,
+ 505,
+ 297
+ ],
+ "score": 1.0,
+ "content": "can model multi-hop GCNs such as Luan et al. (2019); Liao et al. (2019); Abu-El-Haija et al.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 295,
+ 505,
+ 308
+ ],
+ "score": 1.0,
+ "content": "(2019). Higher-order GCs have equivalent expressive power to SoGCs, since they can be reduced to",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 306,
+ 506,
+ 320
+ ],
+ "score": 1.0,
+ "content": "SoGCs as long as coefficient sparsity can be achieved. But this by-product–an uncertain sparsity of",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 317,
+ 506,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 317,
+ 506,
+ 330
+ ],
+ "score": 1.0,
+ "content": "coefficients–is not compatible with gradient-based optimization algorithms. Extensive experiments",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 327,
+ 506,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 327,
+ 506,
+ 340
+ ],
+ "score": 1.0,
+ "content": "(Defferrard et al., 2016) have shown the ineffectiveness of learning higher-order kernels, because",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 338,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 338,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "eigenvalues of graph adjacency matrices diminish when powered. This results in a decreasing nu-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 169,
+ 362
+ ],
+ "score": 1.0,
+ "content": "merical rank of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 170,
+ 349,
+ 199,
+ 362
+ ],
+ "score": 0.92,
+ "content": "{ \\bar { \\mathbf { A } } } ( G ) ^ { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 349,
+ 505,
+ 362
+ ],
+ "score": 1.0,
+ "content": ", which prevent higher-order GCs from aggregating larger-scale information.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 359,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "SoGCs can alleviate this problem by preventing the loss of information due to higher-order powering",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 506,
+ 384
+ ],
+ "score": 1.0,
+ "content": "operation. Finally, higher-order GC lacks nonlinearity. SoGCN can bring a better balance between",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 390,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 390,
+ 396
+ ],
+ "score": 1.0,
+ "content": "the expressive power of low-level layers and nonlinearity among them.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 272,
+ 506,
+ 396
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 409,
+ 408,
+ 422
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 409,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 409,
+ 424
+ ],
+ "score": 1.0,
+ "content": "5 SECOND-ORDER GRAPH CONVOLUTIONAL NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 433,
+ 505,
+ 511
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 504,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 434,
+ 504,
+ 445
+ ],
+ "score": 1.0,
+ "content": "In this section, we introduce other building blocks of GCNs and establish our Second-Order Graph",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 506,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Convolutional Networks (SoGCN) following the fashion of deep learning. First, we promote SoGC",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": "to the multi-channel version analogous to Kipf & Welling (2017). Then we cascade a feature embed-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "ding layer, multiple SoGC layers, and append a readout module. Suppose the multi-channel input is",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 107,
+ 475,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 478,
+ 158,
+ 489
+ ],
+ "score": 0.91,
+ "content": "\\pmb { X } \\in \\mathbb { R } ^ { N \\times D }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 159,
+ 475,
+ 238,
+ 492
+ ],
+ "score": 1.0,
+ "content": "supported in graph",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 238,
+ 478,
+ 267,
+ 489
+ ],
+ "score": 0.9,
+ "content": "G \\in { \\mathcal { G } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 267,
+ 475,
+ 354,
+ 492
+ ],
+ "score": 1.0,
+ "content": ", denote the output of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 479,
+ 359,
+ 488
+ ],
+ "score": 0.68,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 359,
+ 475,
+ 405,
+ 492
+ ],
+ "score": 1.0,
+ "content": "-th layer as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 477,
+ 466,
+ 489
+ ],
+ "score": 0.92,
+ "content": "\\pmb { X } ^ { ( l ) } \\in \\mathbb { R } ^ { N \\times E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 466,
+ 475,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": ", the final",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 103,
+ 485,
+ 507,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 103,
+ 485,
+ 191,
+ 504
+ ],
+ "score": 1.0,
+ "content": "node-level output as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 191,
+ 489,
+ 241,
+ 500
+ ],
+ "score": 0.92,
+ "content": "\\pmb { Y } \\in \\bar { \\mathbb { R } } ^ { N \\times F }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 242,
+ 485,
+ 345,
+ 504
+ ],
+ "score": 1.0,
+ "content": ", or graph-level output as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 488,
+ 382,
+ 500
+ ],
+ "score": 0.91,
+ "content": "\\pmb { Y } \\in \\mathbb { R } ^ { \\pmb { E } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 485,
+ 507,
+ 504
+ ],
+ "score": 1.0,
+ "content": ", we formulate our novel deep",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 241,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 241,
+ 513
+ ],
+ "score": 1.0,
+ "content": "GCN based on SoGC as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 103,
+ 434,
+ 507,
+ 513
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 159,
+ 511,
+ 450,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 159,
+ 511,
+ 450,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 159,
+ 511,
+ 450,
+ 572
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { \\pmb { X } ^ { ( 0 ) } = \\rho \\left( \\pmb { X } ; \\pmb { \\Phi } \\right) , } \\\\ & { \\pmb { X } ^ { ( l + 1 ) } = \\sigma \\left( \\pmb { A } ( G ) ^ { 2 } \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 2 } ^ { ( l + 1 ) } + \\pmb { A } ( G ) \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 1 } ^ { ( l + 1 ) } + \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 0 } ^ { ( l + 1 ) } \\right) , } \\\\ & { \\pmb { Y } = \\tau \\left( \\pmb { X } ^ { ( L ) } ; \\pmb { \\Psi } \\right) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "9780b1edac100fa3f4a86bb987250b3f51c59b8f1a3ff4f05c9d9a0912c717d9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 34,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 159,
+ 511,
+ 450,
+ 531.3333333333334
+ ],
+ "spans": [],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 159,
+ 531.3333333333334,
+ 450,
+ 551.6666666666667
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 159,
+ 551.6666666666667,
+ 450,
+ 572.0000000000001
+ ],
+ "spans": [],
+ "index": 35
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 572,
+ 504,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 572,
+ 135,
+ 590
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 573,
+ 246,
+ 588
+ ],
+ "score": 0.92,
+ "content": "\\Theta _ { i } ^ { ( l ) } \\in \\mathbb { R } ^ { E \\times E } , i = 0 , 1 , 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 572,
+ 409,
+ 590
+ ],
+ "score": 1.0,
+ "content": "are trainable weights for linear filters;",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 410,
+ 575,
+ 504,
+ 587
+ ],
+ "score": 0.89,
+ "content": "\\rho : \\mathbb { R } ^ { N \\times D } \\mathbb { R } ^ { N \\times E }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 104,
+ 586,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 586,
+ 484,
+ 601
+ ],
+ "score": 1.0,
+ "content": "is an equivariant embedder (Maron et al., 2018) with parameters Φ; σ : RN×E → RN×E",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 481,
+ 587,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "is an",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 596,
+ 507,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 596,
+ 281,
+ 613
+ ],
+ "score": 1.0,
+ "content": "activation function. For node-level readout,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 281,
+ 598,
+ 366,
+ 609
+ ],
+ "score": 0.91,
+ "content": "\\tau : \\mathbb { R } ^ { N \\times E } \\stackrel { \\bullet } { \\to } \\mathbb { R } ^ { N \\times F }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 366,
+ 596,
+ 507,
+ 613
+ ],
+ "score": 1.0,
+ "content": "can be a decoder (with parameters",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 107,
+ 608,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 610,
+ 117,
+ 620
+ ],
+ "score": 0.64,
+ "content": "\\Psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 117,
+ 608,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": ") or a nonlinear activation (e.g., softmax) in place of the prior layer. For graph-level output,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 107,
+ 617,
+ 507,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 620,
+ 181,
+ 631
+ ],
+ "score": 0.89,
+ "content": "\\tau : \\mathbb { R } ^ { N \\times E } \\mathbb { R } ^ { E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 617,
+ 507,
+ 635
+ ],
+ "score": 1.0,
+ "content": "should be an invariant readout function (Maron et al., 2018), e.g., channel-wise",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 631,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 631,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "sum, mean or max (Hamilton et al., 2017). In practice, we adopt ReLU as nonlinear activation (i.e.,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 107,
+ 643,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 643,
+ 155,
+ 654
+ ],
+ "score": 0.82,
+ "content": "\\sigma = \\mathrm { R e L U }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 643,
+ 401,
+ 655
+ ],
+ "score": 1.0,
+ "content": "), a multi-layer perceptron (MLP) as the embedding function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 402,
+ 645,
+ 408,
+ 654
+ ],
+ "score": 0.76,
+ "content": "\\rho",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 408,
+ 643,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": ", another MLP for node",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 417,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 224,
+ 666
+ ],
+ "score": 1.0,
+ "content": "regression readout, and sum (",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 224,
+ 654,
+ 238,
+ 664
+ ],
+ "score": 0.38,
+ "content": "\\mathrm { { X u } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 654,
+ 417,
+ 666
+ ],
+ "score": 1.0,
+ "content": "et al., 2019) for graph classification readout.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 39.5,
+ "bbox_fs": [
+ 104,
+ 572,
+ 507,
+ 666
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 678,
+ 243,
+ 689
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 678,
+ 244,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 678,
+ 244,
+ 691
+ ],
+ "score": 1.0,
+ "content": "5.1 GATED RECURRENT UNIT",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 698,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "Gated Recurrent Unit (GRU) has been served as a basic building block in message-passing GNN",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "architectures (Li et al., 2016; Gilmer et al., 2017; Corso et al., 2020). In this subsection, we explore",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 238,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 238,
+ 733
+ ],
+ "score": 1.0,
+ "content": "its application in spectral GCNs.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 46,
+ "bbox_fs": [
+ 106,
+ 698,
+ 505,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 171
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "According to Cho et al. (2014), GRU can utilize gate mechanism to preserve and forget information.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 107
+ ],
+ "score": 1.0,
+ "content": "We hypothesize that a GRU can be trained to remove redundant signals and retain lost features on",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "the spectrum. This function can be used to alleviate the oversmoothing problem of vanilla GCNs",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "by maintaining information from previous layer and canceling the dominance of low-frequencies.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "By the same means, GRU can also relieve the side-effect of ReLU, which is proved to be a special",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 506,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 506,
+ 150
+ ],
+ "score": 1.0,
+ "content": "low-pass filter (Oono & Suzuki, 2019; Cai & Wang, 2020). Even though piled-up SoGCs attain full",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 160
+ ],
+ "score": 1.0,
+ "content": "expressiveness, we show by our experiment that GRU can still facilitate SoGCN in avoiding noises",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 310,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 310,
+ 172
+ ],
+ "score": 1.0,
+ "content": "and enhancing features on the spectrum (Figure 2)",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 175,
+ 505,
+ 231
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 175,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 175,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "Similar to Li et al. (2016); Gilmer et al. (2017), we appends a shared GRU module after each GC",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 187,
+ 505,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 505,
+ 199
+ ],
+ "score": 1.0,
+ "content": "layer, which takes the signal before the GC layer as the hidden state, after the GC layer as the current",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "input. We note that GRU can cooperate with any spectral GCs (Equation 1). When integrated with",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 207,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 207,
+ 505,
+ 224
+ ],
+ "score": 1.0,
+ "content": "SoGCN, we call this special variant SoGCN-GRU. We formulate its implementation by replacing",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 268,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 268,
+ 233
+ ],
+ "score": 1.0,
+ "content": "Equation 10 with Equation 12 as below.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 170,
+ 235,
+ 440,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 170,
+ 235,
+ 440,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 170,
+ 235,
+ 440,
+ 276
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { \\pmb { X } _ { c o n v } ^ { ( l + 1 ) } = \\pmb { A } ( G ) ^ { 2 } \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 2 } ^ { ( l + 1 ) } + \\pmb { A } ( G ) \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 1 } ^ { ( l + 1 ) } + \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 0 } ^ { ( l + 1 ) } , } \\\\ & { \\pmb { X } ^ { ( l + 1 ) } = \\mathrm { G R U } \\left( \\mathrm { R e L U } \\left( \\pmb { X } _ { c o n v } ^ { ( l + 1 ) } \\right) , \\pmb { X } ^ { ( l ) } ; \\pmb { \\Omega } \\right) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "568bdee5e087176294caf31937418aa1f4c4d43a952e75e136d28dcbebb87079.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 170,
+ 235,
+ 440,
+ 248.66666666666666
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 170,
+ 248.66666666666666,
+ 440,
+ 262.3333333333333
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 170,
+ 262.3333333333333,
+ 440,
+ 276.0
+ ],
+ "spans": [],
+ "index": 15
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 281,
+ 487,
+ 294
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 276,
+ 491,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 276,
+ 133,
+ 299
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 280,
+ 164,
+ 294
+ ],
+ "score": 0.92,
+ "content": "X _ { c o n v } ^ { ( l + 1 ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 276,
+ 214,
+ 299
+ ],
+ "score": 1.0,
+ "content": "is the input,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 215,
+ 281,
+ 235,
+ 293
+ ],
+ "score": 0.9,
+ "content": "\\pmb { X } ^ { ( l ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 276,
+ 347,
+ 299
+ ],
+ "score": 1.0,
+ "content": "represents the hidden state,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 283,
+ 357,
+ 292
+ ],
+ "score": 0.81,
+ "content": "\\pmb { \\Omega }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 276,
+ 491,
+ 299
+ ],
+ "score": 1.0,
+ "content": "denotes parameters of the GRU.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 344
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "Figure 2 and Figure 6 verify our conjecture. We observe more steady low-frequency component on",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "the spectrum head and more characteristic bands on the high-frequency tail. Our empirical study in",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "Table 3 also indicates the effectiveness of GRU for spectral GCNs in general. Hence, we suggest",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 333,
+ 428,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 428,
+ 344
+ ],
+ "score": 1.0,
+ "content": "including this recurrent module as another basic building block of our SoGCNs.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 357,
+ 276,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 279,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 279,
+ 370
+ ],
+ "score": 1.0,
+ "content": "5.2 COMPARISON TO RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 488
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 391
+ ],
+ "score": 1.0,
+ "content": "Spectral GCNs. Spectral GCN leverages polynomials in the adjacency matrix to represent graph",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 389,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "convolutional layers (Bruna et al., 2014). Many works have been discussing how to design the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 399,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 399,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "polynomial and choose its degree to compose a localized GC layer. ChebyNet (Defferrard et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 409,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "2016) approximates graph filters using Chebyshev polynomials. Vanilla GCN (Kipf & Welling,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 505,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 505,
+ 434
+ ],
+ "score": 1.0,
+ "content": "2017; Wu et al., 2019) further reduces the GC layer to a degree-one polynomial with first-order and",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 446
+ ],
+ "score": 1.0,
+ "content": "constant terms merged. However, these simplifications cause over-smoothing and performance loss.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Our SoGC incorporates only one hop longer but obtains the full representation power. This design",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 454,
+ 506,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 506,
+ 467
+ ],
+ "score": 1.0,
+ "content": "keeps each layer localized, simple, and easy to implement but makes the whole GCN much more",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 465,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 465,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "powerful. Our work reveals the critical degree of polynomial filters where kernel size is minimized",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 477,
+ 296,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 296,
+ 488
+ ],
+ "score": 1.0,
+ "content": "while filter representation power is maximized.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 499,
+ 505,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 505,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 505,
+ 511
+ ],
+ "score": 1.0,
+ "content": "Multi-Hop GCNs. To exploit multi-scale information, Luan et al. (2019) devises Snowball GCN",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "and Truncated Krylov GCN to capture neighborhoods at different distances. To simulate hop delta",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 521,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "functions, Abu-El-Haija et al. (2019) repeat mixing multi-hop features to identify more topological",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "information. These models exhibit the strength of multi-hop GCNs over one-hop GCNs while leav-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "ing the propagation length as a hyper-parameter. Modeling those multi-hop GCNs as our higher-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "order models, SoGCN possesses the identical representation power but has fixed size and better",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "localization, making SoGC more suitable to be the basic building block in GCNs. It is noteworthy",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "that, although Abu-El-Haija et al. (2019) investigates the two-hop delta function (a Gabor-like filter),",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 601
+ ],
+ "score": 1.0,
+ "content": "their final proposed solution is only a generic class of multi-hop GCNs. The discussion on two-hop",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 598,
+ 312,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 598,
+ 312,
+ 610
+ ],
+ "score": 1.0,
+ "content": "delta functions cannot attain our theoretical results.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 621,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "Expressiveness of GCNs. Most of the works on GCN’s expressiveness are restricted to the over-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "smoothing problem: Li et al. (2018) first poses the over-smoothing problem; Hoang & Maehara",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 643,
+ 505,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 505,
+ 657
+ ],
+ "score": 1.0,
+ "content": "(2019) indicates GCNs are no more than low-pass filters; Luan et al. (2019); Oono & Suzuki (2019)",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "demonstrate the asymptotic behavior of feature activation to a subspace; Cai & Wang (2020) ex-",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "amines the decreasing Dirichlet energy. Unlike them, our established LSS framework can identify",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "specific issues of GCNs with algebraic and geometric interpretations. The over-smoothing prob-",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "lem can be formulated as one of our sub-problems (Section 4.3). In this sense, SoGCN solves",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 464,
+ 712
+ ],
+ "score": 1.0,
+ "content": "a more general expressiveness issue than those relieving over-smoothing problem only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 699,
+ 479,
+ 709
+ ],
+ "score": 0.3,
+ "content": "\\mathrm { { X u } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "2018; Rong et al., 2019; Chen et al., 2020). Ming Chen et al. (2020) introduces identity and initial",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "mapping to recover filter expressiveness. Their analytic framework is also similar to ours. But we",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 46.5
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 306,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 308,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2021",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 171
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "According to Cho et al. (2014), GRU can utilize gate mechanism to preserve and forget information.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 107
+ ],
+ "score": 1.0,
+ "content": "We hypothesize that a GRU can be trained to remove redundant signals and retain lost features on",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "the spectrum. This function can be used to alleviate the oversmoothing problem of vanilla GCNs",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "by maintaining information from previous layer and canceling the dominance of low-frequencies.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 127,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": "By the same means, GRU can also relieve the side-effect of ReLU, which is proved to be a special",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 506,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 506,
+ 150
+ ],
+ "score": 1.0,
+ "content": "low-pass filter (Oono & Suzuki, 2019; Cai & Wang, 2020). Even though piled-up SoGCs attain full",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 160
+ ],
+ "score": 1.0,
+ "content": "expressiveness, we show by our experiment that GRU can still facilitate SoGCN in avoiding noises",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 310,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 310,
+ 172
+ ],
+ "score": 1.0,
+ "content": "and enhancing features on the spectrum (Figure 2)",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 172
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 175,
+ 505,
+ 231
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 175,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 175,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "Similar to Li et al. (2016); Gilmer et al. (2017), we appends a shared GRU module after each GC",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 187,
+ 505,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 505,
+ 199
+ ],
+ "score": 1.0,
+ "content": "layer, which takes the signal before the GC layer as the hidden state, after the GC layer as the current",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "input. We note that GRU can cooperate with any spectral GCs (Equation 1). When integrated with",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 207,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 207,
+ 505,
+ 224
+ ],
+ "score": 1.0,
+ "content": "SoGCN, we call this special variant SoGCN-GRU. We formulate its implementation by replacing",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 268,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 268,
+ 233
+ ],
+ "score": 1.0,
+ "content": "Equation 10 with Equation 12 as below.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 105,
+ 175,
+ 505,
+ 233
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 170,
+ 235,
+ 440,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 170,
+ 235,
+ 440,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 170,
+ 235,
+ 440,
+ 276
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { \\pmb { X } _ { c o n v } ^ { ( l + 1 ) } = \\pmb { A } ( G ) ^ { 2 } \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 2 } ^ { ( l + 1 ) } + \\pmb { A } ( G ) \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 1 } ^ { ( l + 1 ) } + \\pmb { X } ^ { ( l ) } \\pmb { \\Theta } _ { 0 } ^ { ( l + 1 ) } , } \\\\ & { \\pmb { X } ^ { ( l + 1 ) } = \\mathrm { G R U } \\left( \\mathrm { R e L U } \\left( \\pmb { X } _ { c o n v } ^ { ( l + 1 ) } \\right) , \\pmb { X } ^ { ( l ) } ; \\pmb { \\Omega } \\right) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "568bdee5e087176294caf31937418aa1f4c4d43a952e75e136d28dcbebb87079.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 170,
+ 235,
+ 440,
+ 248.66666666666666
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 170,
+ 248.66666666666666,
+ 440,
+ 262.3333333333333
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 170,
+ 262.3333333333333,
+ 440,
+ 276.0
+ ],
+ "spans": [],
+ "index": 15
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 281,
+ 487,
+ 294
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 276,
+ 491,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 276,
+ 133,
+ 299
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 280,
+ 164,
+ 294
+ ],
+ "score": 0.92,
+ "content": "X _ { c o n v } ^ { ( l + 1 ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 276,
+ 214,
+ 299
+ ],
+ "score": 1.0,
+ "content": "is the input,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 215,
+ 281,
+ 235,
+ 293
+ ],
+ "score": 0.9,
+ "content": "\\pmb { X } ^ { ( l ) }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 276,
+ 347,
+ 299
+ ],
+ "score": 1.0,
+ "content": "represents the hidden state,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 283,
+ 357,
+ 292
+ ],
+ "score": 0.81,
+ "content": "\\pmb { \\Omega }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 276,
+ 491,
+ 299
+ ],
+ "score": 1.0,
+ "content": "denotes parameters of the GRU.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 104,
+ 276,
+ 491,
+ 299
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 344
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "Figure 2 and Figure 6 verify our conjecture. We observe more steady low-frequency component on",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "the spectrum head and more characteristic bands on the high-frequency tail. Our empirical study in",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "Table 3 also indicates the effectiveness of GRU for spectral GCNs in general. Hence, we suggest",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 333,
+ 428,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 428,
+ 344
+ ],
+ "score": 1.0,
+ "content": "including this recurrent module as another basic building block of our SoGCNs.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 105,
+ 300,
+ 506,
+ 344
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 357,
+ 276,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 279,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 279,
+ 370
+ ],
+ "score": 1.0,
+ "content": "5.2 COMPARISON TO RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 488
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 391
+ ],
+ "score": 1.0,
+ "content": "Spectral GCNs. Spectral GCN leverages polynomials in the adjacency matrix to represent graph",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 389,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "convolutional layers (Bruna et al., 2014). Many works have been discussing how to design the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 399,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 399,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "polynomial and choose its degree to compose a localized GC layer. ChebyNet (Defferrard et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 409,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "2016) approximates graph filters using Chebyshev polynomials. Vanilla GCN (Kipf & Welling,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 421,
+ 505,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 421,
+ 505,
+ 434
+ ],
+ "score": 1.0,
+ "content": "2017; Wu et al., 2019) further reduces the GC layer to a degree-one polynomial with first-order and",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 446
+ ],
+ "score": 1.0,
+ "content": "constant terms merged. However, these simplifications cause over-smoothing and performance loss.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 457
+ ],
+ "score": 1.0,
+ "content": "Our SoGC incorporates only one hop longer but obtains the full representation power. This design",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 454,
+ 506,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 506,
+ 467
+ ],
+ "score": 1.0,
+ "content": "keeps each layer localized, simple, and easy to implement but makes the whole GCN much more",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 465,
+ 506,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 465,
+ 506,
+ 478
+ ],
+ "score": 1.0,
+ "content": "powerful. Our work reveals the critical degree of polynomial filters where kernel size is minimized",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 477,
+ 296,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 296,
+ 488
+ ],
+ "score": 1.0,
+ "content": "while filter representation power is maximized.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 104,
+ 376,
+ 506,
+ 488
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 499,
+ 505,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 505,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 505,
+ 511
+ ],
+ "score": 1.0,
+ "content": "Multi-Hop GCNs. To exploit multi-scale information, Luan et al. (2019) devises Snowball GCN",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "and Truncated Krylov GCN to capture neighborhoods at different distances. To simulate hop delta",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 521,
+ 505,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 505,
+ 534
+ ],
+ "score": 1.0,
+ "content": "functions, Abu-El-Haija et al. (2019) repeat mixing multi-hop features to identify more topological",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 545
+ ],
+ "score": 1.0,
+ "content": "information. These models exhibit the strength of multi-hop GCNs over one-hop GCNs while leav-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 544,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "ing the propagation length as a hyper-parameter. Modeling those multi-hop GCNs as our higher-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 567
+ ],
+ "score": 1.0,
+ "content": "order models, SoGCN possesses the identical representation power but has fixed size and better",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "localization, making SoGC more suitable to be the basic building block in GCNs. It is noteworthy",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "that, although Abu-El-Haija et al. (2019) investigates the two-hop delta function (a Gabor-like filter),",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 587,
+ 506,
+ 601
+ ],
+ "score": 1.0,
+ "content": "their final proposed solution is only a generic class of multi-hop GCNs. The discussion on two-hop",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 598,
+ 312,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 598,
+ 312,
+ 610
+ ],
+ "score": 1.0,
+ "content": "delta functions cannot attain our theoretical results.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 500,
+ 506,
+ 610
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 621,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "Expressiveness of GCNs. Most of the works on GCN’s expressiveness are restricted to the over-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "smoothing problem: Li et al. (2018) first poses the over-smoothing problem; Hoang & Maehara",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 643,
+ 505,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 505,
+ 657
+ ],
+ "score": 1.0,
+ "content": "(2019) indicates GCNs are no more than low-pass filters; Luan et al. (2019); Oono & Suzuki (2019)",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "demonstrate the asymptotic behavior of feature activation to a subspace; Cai & Wang (2020) ex-",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "amines the decreasing Dirichlet energy. Unlike them, our established LSS framework can identify",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "specific issues of GCNs with algebraic and geometric interpretations. The over-smoothing prob-",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "lem can be formulated as one of our sub-problems (Section 4.3). In this sense, SoGCN solves",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 464,
+ 712
+ ],
+ "score": 1.0,
+ "content": "a more general expressiveness issue than those relieving over-smoothing problem only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 699,
+ 479,
+ 709
+ ],
+ "score": 0.3,
+ "content": "\\mathrm { { X u } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 479,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "2018; Rong et al., 2019; Chen et al., 2020). Ming Chen et al. (2020) introduces identity and initial",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "mapping to recover filter expressiveness. Their analytic framework is also similar to ours. But we",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "generalize their filter space to a graph set, and upper bound its dimension. In the meanwhile, our",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": "SoGCN’s architecture is more lightweight. We investigate the overall expressive power of GCNs",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "by discussing filter completeness. This direction is orthogonal to those studying message-passing",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 493,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 493,
+ 127
+ ],
+ "score": 1.0,
+ "content": "GNNs (Scarselli et al., 2008) and Weisfeiler-Leman GNNs (Xu et al., 2019; Morris et al., 2019).",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 105,
+ 621,
+ 505,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 126
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "generalize their filter space to a graph set, and upper bound its dimension. In the meanwhile, our",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": "SoGCN’s architecture is more lightweight. We investigate the overall expressive power of GCNs",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "by discussing filter completeness. This direction is orthogonal to those studying message-passing",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 493,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 493,
+ 127
+ ],
+ "score": 1.0,
+ "content": "GNNs (Scarselli et al., 2008) and Weisfeiler-Leman GNNs (Xu et al., 2019; Morris et al., 2019).",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 1.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 144,
+ 200,
+ 156
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 201,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 201,
+ 159
+ ],
+ "score": 1.0,
+ "content": "6 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 169,
+ 504,
+ 191
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 169,
+ 505,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 169,
+ 505,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Experiments are conducted on the synthetic dataset in Section 6.1 and on the GNN bench-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 181,
+ 283,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 181,
+ 283,
+ 192
+ ],
+ "score": 1.0,
+ "content": "marks (Dwivedi et al., 2020) in Section 6.2.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 206,
+ 411,
+ 218
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 205,
+ 413,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 205,
+ 413,
+ 219
+ ],
+ "score": 1.0,
+ "content": "6.1 SYNTHETIC GRAPH SPECTRUM DATASET FOR NODE REGRESSION",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 227,
+ 505,
+ 304
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 505,
+ 241
+ ],
+ "score": 1.0,
+ "content": "To validate the expressiveness of SoGCN, and its power to preserve higher-order graph signal, we",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "build a Synthetic Graph Spectrum (SGS) dataset for the node signal filtering regression task. We",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 250,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "construct SGS dataset with random graphs. The learning task is to mimic three types of hand-crafted",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 505,
+ 273
+ ],
+ "score": 1.0,
+ "content": "filtering functions: high-pass, low-pass, and band-pass on the graph spectral space (defined over the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 271,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "graph eigenvectors). There are 1k training graphs, 1k validation graphs, and 2k testing graphs for",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 283,
+ 361,
+ 295
+ ],
+ "score": 1.0,
+ "content": "each filtering function. Each graph is undirected and comprises",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 283,
+ 397,
+ 293
+ ],
+ "score": 0.86,
+ "content": "8 0 \\sim 1 2 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 397,
+ 283,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "nodes. Appendix E covers",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 294,
+ 502,
+ 305
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 294,
+ 502,
+ 305
+ ],
+ "score": 1.0,
+ "content": "more details of our SGS dataset. We choose Mean Absolute Error (MAE) as the evaluation metric.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 318,
+ 503,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 331
+ ],
+ "score": 1.0,
+ "content": "Experimental Setup. We compare SoGCN with vanilla GCN (Kipf & Welling, 2017), first-order",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 329,
+ 327,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 329,
+ 327,
+ 342
+ ],
+ "score": 1.0,
+ "content": "GCN, and higher-order GCNs on the synthetic dataset.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 346,
+ 505,
+ 390
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": "To evaluate each model’s expressiveness purely on the graph kernel design, we remove ReLU acti-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 370
+ ],
+ "score": 1.0,
+ "content": "vations for all tested models. We adopt the Adam optimizer (Kingma & Ba, 2015) in our training",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 505,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 505,
+ 380
+ ],
+ "score": 1.0,
+ "content": "process, with a batch size of 128. The learning rate begins with 0.01 and decays by half once the",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 379,
+ 340,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 379,
+ 340,
+ 391
+ ],
+ "score": 1.0,
+ "content": "validation loss stagnates for more than 10 training epochs.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 132,
+ 450,
+ 479,
+ 559
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 402,
+ 505,
+ 435
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "Table 1: The performance of graph node signal regression with High-Pass, Low-Pass, and Band-",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 413,
+ 505,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 413,
+ 505,
+ 425
+ ],
+ "score": 1.0,
+ "content": "Pass filters (over graph spectral space) as learning target. Each model has 16 GC layers and 16",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 423,
+ 213,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 213,
+ 437
+ ],
+ "score": 1.0,
+ "content": "channels of hidden layers.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 132,
+ 450,
+ 479,
+ 559
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 450,
+ 479,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 450,
+ 479,
+ 559
+ ],
+ "score": 0.985,
+ "html": "